/** * Tests for secrets regex parsing, compilation, and obfuscation. */ import { describe, expect, it, spyOn } from "bun:test"; import * as crypto from "node:crypto"; import * as fs from "node:fs/promises"; import * as os from "node:os"; import * as path from "node:path"; import { type } from "@oh-my-pi/omptype"; import type { AgentMessage } from "@oh-my-pi/pi-agent-core"; import type { AssistantMessage, Context, Message, TextContent } from "@oh-my-pi/pi-ai"; import { builtinCredentialSecretEntries, getExistingSecretPlaceholderKey, getSecretPlaceholderKey, getSecretPlaceholderKeySync, loadSecrets, } from "@oh-my-pi/pi-coding-agent/secrets"; import { deobfuscateAgentMessages, deobfuscateToolArguments, obfuscateMessages, obfuscateProviderContext, obfuscateToolArguments, } from "@oh-my-pi/pi-coding-agent/secrets/message-transform"; import { type SecretEntry, SecretObfuscator } from "@oh-my-pi/pi-coding-agent/secrets/obfuscator"; import { sanitizeSecretFriendlyName, secretEntriesNeedPlaceholderKey, secretEntryNeedsPlaceholderKey, stripPendingSecretPlaceholderSuffix, } from "@oh-my-pi/pi-coding-agent/secrets/placeholder"; import { compileSecretRegex } from "@oh-my-pi/pi-coding-agent/secrets/regex"; import { getActiveProfile, getAgentDir, setProfile } from "@oh-my-pi/pi-utils/dirs"; describe("compileSecretRegex", () => { it("adds global flag when not provided", () => { const regex = compileSecretRegex("api[_-]?key\\s*=\\s*\\w+", "i"); expect(regex.source).toBe("api[_-]?key\\s*=\\s*\\w+"); expect(regex.flags).toBe("gi"); }); it("defaults to global flag when no flags provided", () => { const regex = compileSecretRegex("api[_-]?key\\s*=\\s*\\w+"); expect(regex.source).toBe("api[_-]?key\\s*=\\s*\\w+"); expect(regex.flags).toBe("g"); }); it("rejects invalid regex pattern", () => { expect(() => compileSecretRegex("(")).toThrow(); }); it("rejects invalid regex flags", () => { expect(() => compileSecretRegex("x", "zz")).toThrow(); }); }); describe("builtinCredentialSecretEntries", () => { // Issue #6968: an unconfigured credential-shaped token in a tool result used // to fall through to pi-ai's irreversible `[*_token_redacted]` rewrite, so an // edit-tool `old_string` echoing that placeholder could never match the file. // The contract: the token is hidden from provider-visible text AND restored // byte-exact in tool-call arguments before tool execution. it("hides unconfigured credential-shaped tokens and restores them in tool-call arguments", () => { const obfuscator = new SecretObfuscator(builtinCredentialSecretEntries()); expect(obfuscator.hasSecrets()).toBe(true); const tokens = [`sk-${"a1B-c2D".repeat(7)}e3F`, `ghp_${"aB1".repeat(12)}`, `glpat-${"xY2-".repeat(5)}`]; for (const token of tokens) { const fileLine = `MOONSHOT_API_KEY=${token}`; const providerView = obfuscator.obfuscate(fileLine); expect(providerView).not.toContain(token); // Re-obfuscation is a fixed point: the placeholder itself is never re-matched. expect(obfuscator.obfuscate(providerView)).toBe(providerView); // The model echoes the placeholder into edit-tool old_string verbatim. const args = deobfuscateToolArguments(obfuscator, { old_string: providerView }); expect(args.old_string).toBe(fileLine); } }); }); describe("lazy placeholder key", () => { // The built-in credential entries match dynamically, so they must not force // `secret-placeholder.key` creation for every secrets.enabled session: the // key provider is only invoked when a credential-shaped token is actually // obfuscated, and exactly once across the obfuscator's lifetime. it("resolves the key provider only on the first real credential match", () => { let resolutions = 0; const obfuscator = new SecretObfuscator(builtinCredentialSecretEntries(), () => { resolutions++; return crypto.randomBytes(32).toString("base64url"); }); expect(resolutions).toBe(0); obfuscator.obfuscate("MOONSHOT_API_KEY=huntsville"); expect(resolutions).toBe(0); const token = `sk-${"a1B-c2D".repeat(7)}e3F`; const providerView = obfuscator.obfuscate(`MOONSHOT_API_KEY=${token}`); expect(resolutions).toBe(1); expect(providerView).not.toContain(token); obfuscator.obfuscate(`repeat: ${token}`); expect(resolutions).toBe(1); const args = deobfuscateToolArguments(obfuscator, { old_string: providerView }); expect(args.old_string).toBe(`MOONSHOT_API_KEY=${token}`); }); it("redacts a lazily resolved key from the same input that triggers resolution", () => { const key = "K".repeat(43); const obfuscator = new SecretObfuscator(builtinCredentialSecretEntries(), () => key); const token = `sk-${"a1B-c2D".repeat(7)}e3F`; const providerView = obfuscator.obfuscate(`key=${key}\ncredential=${token}`); expect(providerView).not.toContain(key); expect(providerView).not.toContain(token); }); it("getSecretPlaceholderKeySync creates the key file on demand and shares it with the async readers", async () => { const dir = await fs.mkdtemp(path.join(os.tmpdir(), "omp-lazy-placeholder-key-")); try { expect(await getExistingSecretPlaceholderKey(dir)).toBeUndefined(); const key = getSecretPlaceholderKeySync(dir); expect(key).toMatch(/^[A-Za-z0-9_-]{43}$/); expect(await getExistingSecretPlaceholderKey(dir)).toBe(key); expect(await getSecretPlaceholderKey(dir)).toBe(key); } finally { await fs.rm(dir, { recursive: true, force: true }); } }); }); describe("SecretObfuscator regex behavior", () => { it("obfuscates and deobfuscates regex matches with flags", () => { const obfuscator = new SecretObfuscator([{ type: "regex", content: "api[_-]?key\\s*=\\s*\\w+", flags: "i" }]); const original = "API_KEY=abc and api-key=def"; const obfuscated = obfuscator.obfuscate(original); expect(obfuscated).not.toEqual(original); expect(obfuscator.deobfuscate(obfuscated)).toEqual(original); }); it("supports bare regex patterns without explicit flags", () => { const obfuscator = new SecretObfuscator([{ type: "regex", content: "api[_-]?key\\s*=\\s*\\w+" }]); const text = "api_key=abc and API_KEY=def"; const obfuscated = obfuscator.obfuscate(text); expect(obfuscated).not.toEqual(text); expect(obfuscator.deobfuscate(obfuscated)).toEqual(text); }); it("deobfuscates placeholders through tool-call arguments", () => { const obfuscator = new SecretObfuscator([{ type: "regex", content: "api[_-]?key\\s*=\\s*\\w+", flags: "i" }]); const original = { cmd: "API_KEY=abc and api-key=def", status: "ok", nested: { note: "API_KEY=zzz" } }; const obfuscated = { cmd: obfuscator.obfuscate(original.cmd), status: original.status, nested: { note: obfuscator.obfuscate(original.nested.note) }, }; expect(JSON.stringify(obfuscated)).not.toContain("API_KEY=abc"); expect(deobfuscateToolArguments(obfuscator, obfuscated)).toEqual(original); }); it("obfuscates conversation messages but leaves the system prompt untouched", () => { const secret = "SUPER_SECRET_TOKEN_12345"; const obfuscator = new SecretObfuscator([{ type: "plain", content: secret }]); const context: Context = { systemPrompt: [`workspace contains ${secret}`], messages: [{ role: "user", content: `use ${secret}`, timestamp: 1 }], }; const obfuscated = obfuscateProviderContext(obfuscator, context); // Conversation messages are redacted (and round-trip back to the secret)... expect(JSON.stringify(obfuscated.messages)).not.toContain(secret); expect(obfuscator.deobfuscate(JSON.stringify(obfuscated.messages))).toContain(secret); // ...but the author-controlled system prompt passes through by reference. expect(obfuscated.systemPrompt).toBe(context.systemPrompt); }); it("leaves tool schemas untouched in provider context (no clone, no redaction)", () => { const secret = "SUPER_SECRET_TOKEN_12345"; const obfuscator = new SecretObfuscator([{ type: "plain", content: secret }]); const parameters = type({ note: "string", }).describe(`write ${secret}`); const context: Context = { messages: [], tools: [ { name: "extension_tool", description: `preserve ${secret}`, parameters, }, ], }; const obfuscated = obfuscateProviderContext(obfuscator, context); expect(obfuscated.tools).toBe(context.tools); expect(obfuscated.tools?.[0]?.parameters).toBe(parameters); }); it("redacts user-facing messages and assistant replay content", () => { const secret = "SUPER_SECRET_TOKEN_12345"; const obfuscator = new SecretObfuscator([{ type: "plain", content: secret }]); const userMsg: Message = { role: "user", content: `user says ${secret}`, timestamp: 1 }; const systemDeveloperMsg: Message = { role: "developer", content: `system reminder ${secret}`, timestamp: 1 }; const fileMentionMsg: Message = { role: "developer", content: `${secret}`, attribution: "user", timestamp: 1, }; const assistantMsg: Message = { role: "assistant", content: [ { type: "toolCall", id: "call_1", name: "handoff", arguments: { note: secret }, intent: `handoff ${secret}`, }, ], api: "test", provider: "test", model: "test", usage: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, totalTokens: 0, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 }, }, stopReason: "toolUse", timestamp: 1, }; const toolResultMsg: Message = { role: "toolResult", toolCallId: "call_1", toolName: "read", content: [{ type: "text", text: `tool output ${secret}` }], isError: false, timestamp: 1, }; const obfuscated = obfuscateMessages(obfuscator, [ userMsg, systemDeveloperMsg, fileMentionMsg, assistantMsg, toolResultMsg, ]); // User-facing content and assistant replay fields are redacted. expect(JSON.stringify(obfuscated[0])).not.toContain(secret); expect(JSON.stringify(obfuscated[2])).not.toContain(secret); expect(JSON.stringify(obfuscated[3])).not.toContain(secret); expect(JSON.stringify(obfuscated[4])).not.toContain(secret); // System developer reminders remain untouched. expect(obfuscated[1]).toBe(systemDeveloperMsg); }); it("never rewrites inline image bytes", () => { const secret = "SUPER_SECRET_TOKEN_12345"; const obfuscator = new SecretObfuscator([{ type: "plain", content: secret }]); // A base64 payload that literally contains the secret substring must survive byte-identical; // rewriting it would corrupt the data URL (the Codex "invalid base64" failure). const imageData = `iVBORw0KGgo${secret}AAAASUVORK5CYII=`; const message: Message = { role: "toolResult", toolCallId: "call_1", toolName: "read", content: [ { type: "text", text: `read ${secret}` }, { type: "image", data: imageData, mimeType: "image/png" }, ], isError: false, timestamp: 1, }; const [obfuscated] = obfuscateMessages(obfuscator, [message]) as [typeof message]; const blocks = obfuscated.content; const image = blocks[1]; const text = blocks[0]; // Image bytes untouched... expect(image.type === "image" && image.data).toBe(imageData); // ...while the adjacent text is redacted. expect(text.type === "text" && text.text.includes(secret)).toBe(false); }); it("ignores configured plain secrets shorter than 8 characters", () => { const obfuscator = new SecretObfuscator([{ type: "plain", content: "esp" }]); expect(obfuscator.hasSecrets()).toBe(false); expect(obfuscator.obfuscate("the response despite whitespace")).toBe("the response despite whitespace"); }); it("ignores regex matches shorter than 8 characters", () => { const obfuscator = new SecretObfuscator([{ type: "regex", content: "esp" }]); expect(obfuscator.obfuscate("the response despite whitespace")).toBe("the response despite whitespace"); }); }); describe("getSecretPlaceholderKey", () => { // Isolate under a fresh $HOME (never the real homedir) so mkdir/writeFile // below cannot hit EACCES in a sandboxed CI/review environment where the // real home is read-only, and so the default-arg path under test resolves // through the SAME `getAgentDir()` the runtime uses (see getSecretPlaceholderKey's // docstring) rather than the unrelated `getConfigRootDir()`. async function withTempAgentHome(run: () => Promise): Promise { const originalProfile = getActiveProfile(); const originalHome = process.env.HOME; const tempHomeDir = await fs.mkdtemp(path.join(os.tmpdir(), "omp-secret-key-")); process.env.HOME = tempHomeDir; const homedirSpy = spyOn(os, "homedir").mockReturnValue(tempHomeDir); try { setProfile(undefined); await run(); } finally { setProfile(originalProfile); homedirSpy.mockRestore(); if (originalHome === undefined) { delete process.env.HOME; } else { process.env.HOME = originalHome; } await fs.rm(tempHomeDir, { recursive: true, force: true }); } } it("caches placeholder keys per profile agent dir", async () => { await withTempAgentHome(async () => { const alphaKey = "A".repeat(43); const betaKey = "B".repeat(43); setProfile("alpha"); await fs.mkdir(getAgentDir(), { recursive: true }); await fs.writeFile(path.join(getAgentDir(), "secret-placeholder.key"), alphaKey); expect(await getSecretPlaceholderKey()).toBe(alphaKey); setProfile("beta"); await fs.mkdir(getAgentDir(), { recursive: true }); await fs.writeFile(path.join(getAgentDir(), "secret-placeholder.key"), betaKey); expect(await getSecretPlaceholderKey()).toBe(betaKey); }); }); it("rejects truncated placeholder key files", async () => { await withTempAgentHome(async () => { setProfile("truncated"); await fs.mkdir(getAgentDir(), { recursive: true }); await fs.writeFile(path.join(getAgentDir(), "secret-placeholder.key"), "abc123"); await expect(getSecretPlaceholderKey()).rejects.toThrow("secret placeholder key"); }); }); it("retries empty existing placeholder key files without creating a new one", async () => { await withTempAgentHome(async () => { setProfile("race"); await fs.mkdir(getAgentDir(), { recursive: true }); const keyPath = path.join(getAgentDir(), "secret-placeholder.key"); await fs.writeFile(keyPath, ""); const eventualKey = "C".repeat(43); // Real delay is intentional: this exercises readPlaceholderKeyFile's retry // loop against an actual concurrent filesystem write, not a mockable timer. const writer = Bun.sleep(25).then(() => fs.writeFile(keyPath, eventualKey)); await expect(getExistingSecretPlaceholderKey()).resolves.toBe(eventualKey); await writer; }); }); it("treats an invalid existing placeholder key as absent for redaction", async () => { await withTempAgentHome(async () => { setProfile("invalid-existing"); await fs.mkdir(getAgentDir(), { recursive: true }); await fs.writeFile(path.join(getAgentDir(), "secret-placeholder.key"), "abc123"); // Replace-only/no-secret sessions load the key only to redact it from tool // output; a corrupt key must not block startup, so the existing-key probe // is best-effort. The obfuscate-mode loader still rejects an invalid key. await expect(getExistingSecretPlaceholderKey()).resolves.toBeUndefined(); await expect(getSecretPlaceholderKey()).rejects.toThrow("secret placeholder key"); }); }); }); describe("SecretObfuscator friendlyName placeholders", () => { it("prefixes plain secret placeholders with sanitized friendly names", () => { const secret = "github_pat_abc123"; const obfuscator = new SecretObfuscator([{ type: "plain", content: secret, friendlyName: "GitHub Token!" }]); const input = `use ${secret} now`; const obfuscated = obfuscator.obfuscate(input); expect(obfuscated).toMatch(/\$\$GITHUBTOKEN_[A-Z0-9]+:L\$\$/); expect(obfuscated).not.toContain(secret); expect(obfuscator.deobfuscate(obfuscated)).toBe(input); }); it("rejects a friendlyName that is a case/punctuation variant of its own secret, but still applies unrelated friendly names", () => { // `#friendlyNameCollidesWithSecret` used to compare the sanitized (uppercased, // alnum-only) friendly name against each secret's RAW value, so a friendlyName // that was merely a case/punctuation variant of its own secret (e.g. // "GitHub_Pat_Abc123" labeling "github_pat_abc123") slipped through: sanitizing // the label produced "GITHUBPATABC123", which never literally appears inside the // lowercase, underscored raw secret string. The fix sanitizes the secret value // the same way before comparing, so this exact-content-under-normalization case // is now caught and the secret falls back to a bare placeholder — while a // genuinely unrelated friendly name is untouched and still gets its prefix. const collidingSecret = "github_pat_abc123"; const collidingObfuscator = new SecretObfuscator([ { type: "plain", content: collidingSecret, friendlyName: "GitHub_Pat_Abc123" }, ]); const collidingObfuscated = collidingObfuscator.obfuscate(collidingSecret); expect(collidingObfuscated).not.toMatch(/GITHUBPATABC123_/); expect(collidingObfuscated).toMatch(/^\$\$[A-Z0-9]+:L\$\$$/); expect(collidingObfuscator.deobfuscate(collidingObfuscated)).toBe(collidingSecret); const distinctSecret = "github_pat_xyz789"; const distinctObfuscator = new SecretObfuscator([ { type: "plain", content: distinctSecret, friendlyName: "GitHub Token" }, ]); const distinctObfuscated = distinctObfuscator.obfuscate(distinctSecret); expect(distinctObfuscated).toMatch(/^\$\$GITHUBTOKEN_[A-Z0-9]+:L\$\$$/); expect(distinctObfuscator.deobfuscate(distinctObfuscated)).toBe(distinctSecret); }); it("rejects a friendlyName equal to its own secret even when the secret's sanitized form exceeds the 32-char display cap", () => { // `#friendlyNameCollidesWithSecret` used to compare the friendly name // against each secret's sanitized value using the ALREADY-CAPPED, // display-length friendly name (sliced to `MAX_FRIENDLY_NAME_LEN` = 32 // chars by `sanitizeSecretFriendlyName`) rather than the full, // un-truncated sanitized form. For a secret whose sanitized (alnum-only, // uppercased) form is longer than 32 characters, the 32-char-capped // label being checked could never `.includes()` the longer sanitized // secret, so a friendlyName set to the secret's own value slipped past // the collision guard entirely — and the secret's first 32 sanitized // characters were accepted and baked into the placeholder as a // visible prefix (e.g. "#GITHUBPATABCDEFGHIJKLMNOPQRSTUV_:L#"), // leaking part of the secret. The check now runs against the full, // un-truncated sanitized label before the 32-char cap is applied for // display, so secrets longer than the cap are still fully compared // and caught. const longSecret = "github_pat_abcdefghijklmnopqrstuvwxyz0123456789"; const obfuscator = new SecretObfuscator([{ type: "plain", content: longSecret, friendlyName: longSecret }]); const obfuscated = obfuscator.obfuscate(longSecret); expect(obfuscated).not.toMatch(/GITHUBPAT/); expect(obfuscated).toMatch(/^\$\$[A-Z0-9]+:L\$\$$/); expect(obfuscator.deobfuscate(obfuscated)).toBe(longSecret); }); it("rejects a friendlyName whose sanitized value is exactly the first 32 characters of a longer secret", () => { // Regression: `sanitizedLabelCollidesWithSecret` must also reject a // friendlyName that is a strict PREFIX of a longer secret's sanitized // form, not just one that equals the secret outright (the case above). // A friendlyName that sanitizes to exactly `MAX_FRIENDLY_NAME_LEN` (32) // characters and happens to be the secret's own leading 32 sanitized // characters must still be treated as a collision — otherwise those 32 // characters are accepted and baked into the placeholder verbatim as a // visible, secret-derived prefix (e.g. // "#GITHUBPATABCDEFGHIJKLMNOPQRSTUVW_:L#"). const longSecret = "github_pat_abcdefghijklmnopqrstuvwxyz0123456789"; const leakedPrefix = "GITHUBPATABCDEFGHIJKLMNOPQRSTUVW"; // first 32 sanitized chars of longSecret const obfuscator = new SecretObfuscator([{ type: "plain", content: longSecret, friendlyName: leakedPrefix }]); const obfuscated = obfuscator.obfuscate(longSecret); expect(obfuscated).not.toContain(leakedPrefix); expect(obfuscated).toMatch(/^\$\$[A-Z0-9]+:L\$\$$/); expect(obfuscator.deobfuscate(obfuscated)).toBe(longSecret); }); it("uses regex entry friendly names for discovered matches", () => { const secret = "tok_abc123"; const obfuscator = new SecretObfuscator([{ type: "regex", content: "tok_[a-z0-9]+", friendlyName: "API Key" }]); const input = `use ${secret} please`; const obfuscated = obfuscator.obfuscate(input); expect(obfuscated).toMatch(/\$\$APIKEY_[A-Z0-9]+:L\$\$/); expect(obfuscated).not.toContain(secret); expect(obfuscator.deobfuscate(obfuscated)).toBe(input); }); it("rejects a regex entry friendlyName that is itself a live match for its own pattern", () => { // `#friendlyNameCollidesWithSecret` used to run each regex entry's pattern // against the SANITIZED friendly name (uppercased, non-alphanumeric // stripped) instead of the raw one. A case-sensitive/punctuated pattern // like `tok_[a-z0-9]+` requires a literal lowercase underscore, so it could // never match the sanitized label "TOKABC123" even though the raw // friendlyName "tok_abc123" is itself a live match for that very pattern. // That let a secret-shaped friendlyName slip through and get stamped // (minus separators, uppercased) into every placeholder minted for it. The // check now runs the pattern against the raw, pre-sanitization // friendlyName, so this case is caught and the secret falls back to a bare // placeholder. const secret = "tok_abc123"; const obfuscator = new SecretObfuscator([ { type: "regex", content: "tok_[a-z0-9]+", friendlyName: "tok_abc123" }, ]); const input = `use ${secret} now`; const obfuscated = obfuscator.obfuscate(input); expect(obfuscated).not.toMatch(/TOKABC123_/); expect(obfuscated).toMatch(/^use \$\$[A-Z0-9]+:L\$\$ now$/); expect(obfuscator.deobfuscate(obfuscated)).toBe(input); }); it("rejects a regex entry friendlyName that is the normalized form of a value the regex discovers", () => { // `#friendlyNameCollidesWithSecret` also used to test the RAW friendlyName // directly against the regex entry's own pattern. That misses labels which // are already normalized (uppercased, separators stripped): "TOKABC123" // has no lowercase letters or underscore, so it can never match the // case-sensitive, punctuated pattern `tok_[a-z0-9]+` directly - even though // "TOKABC123" is exactly the normalized rendering of "tok_abc123", a value // that pattern actually discovers. Nothing compared the label against the // literal match either, so this friendlyName slipped through and got // stamped into the placeholder. The check now also compares the sanitized // label against the sanitized value of the current secret being minted, // catching this on the secret's very first mint. const obfuscator = new SecretObfuscator([{ type: "regex", content: "tok_[a-z0-9]+", friendlyName: "TOKABC123" }]); const input = "use tok_abc123 now"; const obfuscated = obfuscator.obfuscate(input); expect(obfuscated).not.toMatch(/TOKABC123_/); expect(obfuscated).toMatch(/^use \$\$[A-Z0-9]+:L\$\$ now$/); expect(obfuscator.deobfuscate(obfuscated)).toBe(input); }); it("omits an unrelated regex entry's friendly label that normalizes to a value a later regex entry discovers in the same input", () => { // Earlier coverage only caught a regex entry's friendlyName colliding with // its own pattern/value. Regex placeholders are minted lazily in entries[] // order, so an unrelated entry listed first could stamp a normalized form // of a later regex match into its placeholder before that later entry // discovered the raw value in the same input. The upfront input scan keeps // labels from exposing any regex-protected value present in this pass, // regardless of entry order. const obfuscator = new SecretObfuscator([ { type: "regex", content: "zeta_[a-z0-9]+", friendlyName: "TOKABC123" }, { type: "regex", content: "tok_[a-z0-9]+" }, ]); const input = "use zeta_secret1 and tok_abc123 now"; const obfuscated = obfuscator.obfuscate(input); expect(obfuscated).not.toContain("TOKABC123_"); expect(obfuscated).not.toContain("zeta_secret1"); expect(obfuscated).not.toContain("tok_abc123"); expect(obfuscated).toMatch(/^use \$\$[A-Z0-9]{4,}(?::[ULCM])?\$\$ and \$\$[A-Z0-9]{4,}(?::[ULCM])?\$\$ now$/); expect(obfuscator.deobfuscate(obfuscated)).toBe(input); }); it("omits an unrelated regex entry's friendly label that normalizes to a regex-protected value produced by a plain replace mapping", () => { // Regression: the upfront regex-secret scan that seeds the friendly-label // collision set only ran once, against the ORIGINAL input, before plain // replace-mode mappings execute. A replace-mode plain secret can emit // text a later regex entry protects (here `X` -> `tok_abc123`, discovered // by `tok_[a-z0-9]+`) even though that exact text never appeared in the // raw input. Without also re-scanning the POST-replace text, an unrelated // regex entry's friendlyName that normalizes to that produced value still // got stamped into its placeholder. The obfuscator now merges regex // matches collected after the replace phase into the same collision set. const obfuscator = new SecretObfuscator([ { type: "plain", content: "X", mode: "replace", replacement: "tok_abc123" }, { type: "regex", content: "zeta_[a-z0-9]+", friendlyName: "TOKABC123" }, { type: "regex", content: "tok_[a-z0-9]+" }, ]); const input = "use X and zeta_secret1 now"; const obfuscated = obfuscator.obfuscate(input); expect(obfuscated).not.toContain("TOKABC123_"); expect(obfuscated).not.toContain("zeta_secret1"); expect(obfuscated).not.toContain("tok_abc123"); expect(obfuscated).toMatch(/^use \$\$[A-Z0-9]{4,}(?::[ULCM])?\$\$ and \$\$[A-Z0-9]{4,}(?::[ULCM])?\$\$ now$/); // Deobfuscation restores the two obfuscate-mode (regex-discovered) // placeholders to the values that were actually matched — the replace- // produced `tok_abc123` and the raw `zeta_secret1` — but the one-way // replace mapping never restores `X` itself. expect(obfuscator.deobfuscate(obfuscated)).toBe("use tok_abc123 and zeta_secret1 now"); }); it("omits an unrelated regex entry's friendly label that normalizes to a regex-protected value produced by a regex replace mapping", () => { // Regression: same upfront-collision-set gap as the plain-replace case // above, but the value-producing entry is itself a REGEX replace mapping // (`X` -> `tok_abc123` via `{ type: "regex", mode: "replace" }`) rather // than a plain replace secret. The regex-entries loop processes entries // in order and must merge newly discovered regex-protected values into // the friendly-label collision set after EACH replace-mode regex entry // runs, so the unrelated `zeta_[a-z0-9]+` entry's friendlyName // "TOKABC123" must still be rejected even though `tok_abc123` never // appeared in the raw input — it only exists because the regex replace // entry ahead of it produced it. const obfuscator = new SecretObfuscator([ { type: "regex", mode: "replace", content: "X", replacement: "tok_abc123" }, { type: "regex", content: "zeta_[a-z0-9]+", friendlyName: "TOKABC123" }, { type: "regex", content: "tok_[a-z0-9]+" }, ]); const input = "use X and zeta_secret1 now"; const obfuscated = obfuscator.obfuscate(input); expect(obfuscated).not.toContain("TOKABC123_"); expect(obfuscated).not.toContain("zeta_secret1"); expect(obfuscated).not.toContain("tok_abc123"); expect(obfuscated).toMatch(/^use \$\$[A-Z0-9]{4,}(?::[ULCM])?\$\$ and \$\$[A-Z0-9]{4,}(?::[ULCM])?\$\$ now$/); // Deobfuscation restores the two obfuscate-mode (regex-discovered) // placeholders to the values that were actually matched — the // regex-replace-produced `tok_abc123` and the raw `zeta_secret1` — but // the one-way regex replace mapping never restores the original `X`. expect(obfuscator.deobfuscate(obfuscated)).toBe("use tok_abc123 and zeta_secret1 now"); }); it("omits a plain secret's own friendly label that normalizes to a regex-protected value produced by a later regex replace mapping", () => { // Regression: the two collision-set gaps above cover an UNRELATED regex // entry's friendlyName colliding with a later-produced regex-protected // value. This covers a PLAIN obfuscate-mode secret's OWN friendlyName. // Plain obfuscate secrets mint their placeholder in the CONSTRUCTOR — // before `obfuscate()` ever runs and forecasts post-replace regex // values into `#currentRegexSecretValues` — so a friendlyName that // merely normalizes to a value a LATER regex replace mapping produces // (here `X` -> `tok_abc123`, protected by `tok_[a-z0-9]+`) can get // baked into the placeholder at mint time even though `tok_abc123` // never appears in the raw input. The obfuscator must re-check the // baked-in friendly prefix against the freshly forecasted collision // set on every `obfuscate()` call and fall back to a bare placeholder // when it collides, so the friendly prefix never exposes the // normalized future regex secret. const obfuscator = new SecretObfuscator([ { type: "plain", content: "OTHERSECRET", friendlyName: "TOKABC123" }, { type: "regex", mode: "replace", content: "X", replacement: "tok_abc123" }, { type: "regex", content: "tok_[a-z0-9]+" }, ]); const input = "use OTHERSECRET and X now"; const obfuscated = obfuscator.obfuscate(input); expect(obfuscated).not.toContain("TOKABC123_"); expect(obfuscated).not.toContain("OTHERSECRET"); expect(obfuscated).not.toContain("tok_abc123"); expect(obfuscated).toMatch(/^use \$\$[A-Z0-9]{4,}:U\$\$ and \$\$[A-Z0-9]{4,}:L\$\$ now$/); // Deobfuscation restores the plain secret's placeholder to `OTHERSECRET` // and the regex-discovered placeholder to the value actually matched // (`tok_abc123`); the one-way regex replace mapping never restores `X`. expect(obfuscator.deobfuscate(obfuscated)).toBe("use OTHERSECRET and tok_abc123 now"); }); it("omits a plain secret's own friendly label that normalizes to a regex-protected value produced through a DEFAULT (no custom replacement) regex replace hop", () => { // Regression: the forecast above only ever simulated CUSTOM-replacement // regex replace entries (`entry.replacement !== undefined`). A DEFAULT // replace-mode entry (no `replacement` configured) still produces // deterministic output — `X` (one byte) always redacts to `Z` — and that // output can itself be the trigger a LATER replace entry needs to fire. // Here `X` -> `Z` is the default hop, and only `Z` -> `tok_abc123` (a // second, custom-replacement entry) actually produces the // `tok_[a-z0-9]+`-protected value. Skipping the default hop means the // forecast never sees `Z` in the simulated text, so the second entry // never matches and `tok_abc123` is never simulated — the plain secret's // own friendlyName "TOKABC123" (which normalizes to that same value) // wrongly survives into its placeholder. The forecast must simulate the // default hop's own deterministic output too, not just custom // replacements, before checking friendly labels. const obfuscator = new SecretObfuscator([ { type: "plain", content: "OTHERSECRET", friendlyName: "TOKABC123" }, { type: "regex", mode: "replace", content: "X" }, { type: "regex", mode: "replace", content: "Z", replacement: "tok_abc123" }, { type: "regex", content: "tok_[a-z0-9]+" }, ]); const input = "use OTHERSECRET and X now"; const obfuscated = obfuscator.obfuscate(input); expect(obfuscated).not.toContain("TOKABC123_"); expect(obfuscated).not.toContain("OTHERSECRET"); expect(obfuscated).not.toContain("tok_abc123"); expect(obfuscated).toMatch(/^use \$\$[A-Z0-9]{4,}:U\$\$ and \$\$[A-Z0-9]{4,}:L\$\$ now$/); // Deobfuscation restores the plain secret's placeholder to `OTHERSECRET` // and the regex-discovered placeholder to the value actually matched // (`tok_abc123`); the default `X` -> `Z` hop is one-way, so neither `X` // nor the intermediate `Z` marker is ever restored. expect(obfuscator.deobfuscate(obfuscated)).toBe("use OTHERSECRET and tok_abc123 now"); }); it("strips an unsafe prefix from an unknown historical friendly placeholder", () => { const obfuscator = new SecretObfuscator([ { type: "plain", content: "OTHERSECRET", friendlyName: "TOKABC123" }, { type: "regex", content: "tok_[a-z0-9]+" }, ]); const stalePlaceholder = "$$TOKABC123_OLDHASH:L$$"; expect(obfuscator.stripUnsafeFriendlyPlaceholderPrefixes(stalePlaceholder, new Set(["tok_abc123"]))).toBe( "$$OLDHASH:L$$", ); }); it("strips unsafe prefixes from overlapping historical placeholders", () => { const obfuscator = new SecretObfuscator([ { type: "plain", content: "OTHERSECRET", friendlyName: "TOKABC123" }, { type: "regex", content: "tok_[a-z0-9]+" }, ]); const stalePlaceholders = "$$FOOO$$$$TOKABC123_OLDHASH:L$$"; expect(obfuscator.stripUnsafeFriendlyPlaceholderPrefixes(stalePlaceholders, new Set(["tok_abc123"]))).toBe( "$$FOOO$$$$OLDHASH:L$$", ); }); it("strips an already-obfuscated placeholder's unsafe friendly prefix carried in from earlier history when a later input reveals the regex-protected value it normalizes to", () => { // Regression: the two same-call forecasts above only re-check a friendly // prefix at the moment its placeholder is FIRST substituted into THIS // call's output, against a collision set forecasted from THIS call's own // input. A friendly placeholder minted in an EARLIER call — before // `tok_abc123` had ever appeared anywhere — is baked into provider-visible // history and re-enters a LATER `obfuscate()` call verbatim as part of the // input text (the SDK re-obfuscates the whole message array every turn). // The obfuscator must re-check that ALREADY-PRESENT placeholder's baked-in // prefix against the freshly forecasted collision set on every call too — // not just newly minted placeholders — and fall back to the // friendly-name-independent bare alias once a later turn's regex-protected // value normalizes to that prefix. Copying the placeholder verbatim would // leave "TOKABC123_" standing in for `tok_abc123`, the very value the // `tok_[a-z0-9]+` regex is configured to hide. const obfuscator = new SecretObfuscator([ { type: "plain", content: "OTHERSECRET", friendlyName: "TOKABC123" }, { type: "regex", content: "tok_[a-z0-9]+" }, ]); // Turn 1: tok_abc123 has not appeared anywhere yet, so "TOKABC123" is not // a collision — the friendly prefix is applied and persisted into history. const turn1 = obfuscator.obfuscate("use OTHERSECRET now"); expect(turn1).toMatch(/^use \$\$TOKABC123_[A-Z0-9]+:U\$\$ now$/); const oldPlaceholder = turn1.match(/\$\$TOKABC123_[A-Z0-9]+:U\$\$/)![0]; const bareAlias = oldPlaceholder.replace(/^\$\$TOKABC123_/, "$$"); // Turn 2: the already-obfuscated turn-1 output re-enters as prior history // alongside NEW text that reveals tok_abc123 — a regex-protected value that // normalizes to exactly the prefix already baked into the turn-1 placeholder. const turn2Input = `${turn1} and now use tok_abc123`; const turn2 = obfuscator.obfuscate(turn2Input); expect(turn2).not.toContain("TOKABC123_"); expect(turn2).not.toContain(oldPlaceholder); expect(turn2).not.toContain("tok_abc123"); // The preserved placeholder is re-emitted as its bare, friendly-name-independent alias. expect(turn2).toContain(bareAlias); // OTHERSECRET still round-trips via the bare alias, and the newly // discovered regex secret restores to the value actually matched. expect(obfuscator.deobfuscate(turn2)).toBe("use OTHERSECRET now and now use tok_abc123"); // Stripping the unsafe prefix is itself a fixed point: a further pass must // not resurrect the friendly prefix or otherwise change the bytes. expect(obfuscator.obfuscate(turn2)).toBe(turn2); }); it("does not replace plain secrets inside generated friendly placeholders", () => { const longSecret = "long-secret-token"; const prefixSecret = "TOKENABC"; const obfuscator = new SecretObfuscator([ { type: "plain", content: longSecret, friendlyName: "token" }, { type: "plain", content: prefixSecret }, ]); const input = `${longSecret} ${prefixSecret}`; const obfuscated = obfuscator.obfuscate(input); expect(obfuscated).toMatch(/^\$\$TOKEN_[A-Z0-9]+:L\$\$ \$\$[A-Z0-9]+:U\$\$$/); expect(obfuscated).not.toContain(longSecret); expect(obfuscator.deobfuscate(obfuscated)).toBe(input); }); it("redacts configured secrets that already look like placeholders", () => { const secret = "#PASSWORD123#"; const obfuscator = new SecretObfuscator([{ type: "plain", content: secret }]); const obfuscated = obfuscator.obfuscate(`value ${secret}`); expect(obfuscated).not.toContain(secret); expect(obfuscated).toMatch(/^value \$\$[A-Z0-9]+:U\$\$$/); expect(obfuscator.deobfuscate(obfuscated)).toBe(`value ${secret}`); }); it("redacts regex matches after rejected placeholder-shaped text", () => { const obfuscator = new SecretObfuscator([ { type: "plain", content: "ABCDEFGH" }, { type: "regex", content: "SECRET[A-Z]{12}", mode: "replace" }, ]); const obfuscated = obfuscator.obfuscate("#SECRETUVKABCDEFGHI"); expect(obfuscated).not.toContain("SECRET"); expect(obfuscated).not.toContain("ABCDEFGH"); }); it("redacts regex matches that span known placeholders", () => { const obfuscator = new SecretObfuscator([ { type: "plain", content: "abcdefgh" }, { type: "regex", content: "api_key=\\S+", friendlyName: "api-key" }, ]); const obfuscated = obfuscator.obfuscate("api_key=abcdefghXYZ"); expect(obfuscated).toMatch(/^\$\$APIKEY_[A-Z0-9]+:M\$\$$/); expect(obfuscated).not.toContain("abcdefgh"); expect(obfuscator.deobfuscate(obfuscated)).toBe("api_key=abcdefghXYZ"); }); it("redacts bounded obfuscate regex spans around generated placeholders", () => { const obfuscator = new SecretObfuscator([ { type: "plain", content: "abcdefgh" }, { type: "regex", content: "api_key=[A-Za-z0-9]{11}", friendlyName: "api-key" }, ]); const obfuscated = obfuscator.obfuscate("api_key=abcdefghXYZ"); expect(obfuscated).toMatch(/^\$\$APIKEY_[A-Z0-9]+:M\$\$$/); expect(obfuscated).not.toContain("abcdefgh"); expect(obfuscator.deobfuscate(obfuscated)).toBe("api_key=abcdefghXYZ"); }); it("obfuscates bounded regex remainders around prior placeholders", () => { const obfuscator = new SecretObfuscator([ { type: "plain", content: "abcdefgh" }, { type: "regex", content: "api_key=[A-Za-z0-9]{11}", friendlyName: "api-key" }, ]); const token = obfuscator.obfuscate("abcdefgh"); const obfuscated = obfuscator.obfuscate(`api_key=${token}XYZ`); expect(obfuscated).not.toContain("api_key="); expect(obfuscated).not.toContain("XYZ"); expect(obfuscated).toContain(token); expect(obfuscator.deobfuscate(obfuscated)).toBe("api_key=abcdefghXYZ"); expect(obfuscator.obfuscate(obfuscated)).toBe(obfuscated); }); it("skips sub-threshold obfuscate regex matches that straddle a generated placeholder", () => { // `[A-Z]{6}` only ever matches 6 chars — below MIN_OBFUSCATE_SECRET_LEN — but // when a match overlaps the plain secret's placeholder its range is extended // across the whole `#…#` token. Guarding on that rewritten span (instead of // the regex's own match length) let the short match re-placeholder across the // token and corrupt round-trip deobfuscation (e.g. "XXSECRETUVYY" dropping to // "XXSECRETUV"). The plain secret must stay hidden, the surrounding literals // must survive, and the round trip must be exact. const obfuscator = new SecretObfuscator([ { type: "plain", content: "SECRETUV" }, { type: "regex", content: "[A-Z]{6}" }, ]); const obfuscated = obfuscator.obfuscate("XXSECRETUVYY"); expect(obfuscated).not.toContain("SECRETUV"); expect(obfuscated.startsWith("XX")).toBe(true); expect(obfuscated.endsWith("YY")).toBe(true); expect(obfuscator.deobfuscate(obfuscated)).toBe("XXSECRETUVYY"); // Re-obfuscation MUST be a fixed point: a sub-threshold match straddling the // now-input placeholder must not rewrite the surrounding context into fresh // placeholders, or provider-visible history and prompt-cache prefixes drift. expect(obfuscator.obfuscate(obfuscated)).toBe(obfuscated); }); it("leaves placeholders intact when a regex match cuts across their expanded value", () => { // `[A-Z]{8}` meets MIN_OBFUSCATE_SECRET_LEN, but on "YYBBABCDEFGHSECRETUV" // (plain secret `ABCDEFGH`) its greedy 8-char matches start/end inside the // secret's placeholder expansion ("YYBBABCD", "EFGHSECR"). Snapping each to // the whole `#…#` token mapped them to overlapping ranges that clobbered on // apply and dropped the `SECR` bytes (round-trip restored "YYBBABCDEFGHETUV"). // The scan now resumes past the cut placeholder instead, so the secret stays // hidden, no bytes are lost, the trailing 8-char run is obfuscated on its own, // and re-obfuscation is a fixed point. const obfuscator = new SecretObfuscator([ { type: "plain", content: "ABCDEFGH" }, { type: "regex", content: "[A-Z]{8}" }, ]); const obfuscated = obfuscator.obfuscate("YYBBABCDEFGHSECRETUV"); expect(obfuscated).not.toContain("ABCDEFGH"); expect(obfuscated).not.toContain("SECRETUV"); expect(obfuscator.deobfuscate(obfuscated)).toBe("YYBBABCDEFGHSECRETUV"); expect(obfuscator.obfuscate(obfuscated)).toBe(obfuscated); }); it("keeps the full cut-match length when a regex match's outside prefix is clamped below the short-match floor", () => { // Regression: when an obfuscate-mode regex match starts before an // already-generated placeholder and gets clamped down to just the // wholly-outside prefix portion before that placeholder, the short-match // guard (MIN_OBFUSCATE_SECRET_LEN) must measure the FULL original match // length in the expanded scan view, not the shorter clamped-prefix length. // Reusing the sibling "[A-Z]{8}" straddle above but split across two // obfuscate() calls (same instance/key, so the first call's placeholder is // recognized as already-generated on the second): call 1 mints ABCDEFGH's // placeholder alone; call 2 re-scans "YYBB" immediately followed by that // placeholder, and the SAME "[A-Z]{8}" regex re-matches across "YYBB" plus // the placeholder's expanded value, clamping to just the 4-byte "YYBB" // prefix. Before the fix, that clamped 4-byte length (not the true 8-byte // full match) was checked against the floor, so "YYBB" was wrongly skipped // as too short and left raw/unredacted in the output. const obfuscator = new SecretObfuscator([ { type: "plain", content: "ABCDEFGH" }, { type: "regex", content: "[A-Z]{8}" }, ]); const placeholder = obfuscator.obfuscate("ABCDEFGH"); expect(placeholder).not.toContain("ABCDEFGH"); const second = obfuscator.obfuscate(`YYBB${placeholder}`); expect(second).not.toContain("YYBB"); expect(obfuscator.deobfuscate(second)).toBe("YYBBABCDEFGH"); }); it("keeps replace regexes from rewriting placeholders their match cuts across", () => { // Same partial-placeholder cut as above but in replace mode: `[A-Z]{8}` // matches straddle the `ABCDEFGH` placeholder. Redacting only the bytes // outside the snapped token was not a fixed point — the deterministic scramble // of the prefix drifted across re-obfuscation passes ("ZZgK…" → "ZZgZ…"). The // scan now resumes past the cut placeholder, so the cut secret stays hidden as // its placeholder, the trailing 8-char run is redacted on its own, and // re-obfuscation is stable. const obfuscator = new SecretObfuscator([ { type: "plain", content: "ABCDEFGH" }, { type: "regex", content: "[A-Z]{8}", mode: "replace" }, ]); const obfuscated = obfuscator.obfuscate("YYBBABCDEFGHSECRETUV"); expect(obfuscated).not.toContain("ABCDEFGH"); expect(obfuscated).not.toContain("SECRETUV"); expect(obfuscator.obfuscate(obfuscated)).toBe(obfuscated); }); it("replace-mode re-obfuscation is a fixed point when a spillover match straddles a prior-call placeholder whose value satisfies the regex", () => { // `[A-Z0-9]{8,12}` in replace mode greedily spans `SECRETUV` into the // `ABCDEFGH` placeholder on re-obfuscation — `ABCDEFGH` alone satisfies the // regex. Previously the short non-matching spillover bytes abutting the // placeholder were fed into the deterministic scramble on every pass // (e.g. `…#…#ZZJ5sotJ` → `…#…#ZZpvsotJ`), invalidating provider prompt-cache // prefixes despite no new input. The fix: when a replace match straddles a // placeholder whose own deobfuscated value satisfies the regex, the surrounding // spillover bytes are left verbatim instead of being re-scrambled. const obf = new SecretObfuscator( [ { type: "plain", content: "ABCDEFGH" }, { type: "regex", content: "[A-Z0-9]{8,12}", mode: "replace" }, ], "Q".repeat(43), ); const first = obf.obfuscate("YYBBABCDEFGHSECRETUV"); // Core regression: re-obfuscation must be a fixed point. expect(obf.obfuscate(first)).toBe(first); // Stable across multiple passes. expect(obf.obfuscate(obf.obfuscate(first))).toBe(first); // Plain secret must not leak into provider-visible output. expect(first).not.toContain("ABCDEFGH"); }); it("redacts an independently matching prefix before a cut placeholder", () => { // A regex match that starts in outside text and ends inside a generated // placeholder's expanded value is not rewritten across the token (that drops // bytes / drifts the redaction). But when the wholly-outside prefix before the // placeholder is itself a complete match it is provider-visible secret-shaped // content, not a drift artifact. Regression: `[A-Z0-9]{8,12}` greedily // spanning `SECRETUV` into an `ABCDEFGH` placeholder returned `SECRETUV#…#`, // leaving `SECRETUV` visible even though it independently satisfies the regex. // Obfuscate mode: the prefix gets its own reversible placeholder and the whole // input still round-trips. const obf = new SecretObfuscator( [ { type: "plain", content: "ABCDEFGH" }, { type: "regex", content: "[A-Z0-9]{8,12}" }, ], "Q".repeat(43), ); const out = obf.obfuscate("SECRETUVABCDEFGH"); expect(out).not.toContain("SECRETUV"); expect(out).not.toContain("ABCDEFGH"); expect(obf.deobfuscate(out)).toBe("SECRETUVABCDEFGH"); expect(obf.obfuscate(out)).toBe(out); // Replace mode: the prefix is redacted one-way while the cut secret's // placeholder is preserved and still restores. const repl = new SecretObfuscator( [ { type: "plain", content: "ABCDEFGH" }, { type: "regex", content: "[A-Z0-9]{8,12}", mode: "replace" }, ], "Q".repeat(43), ); const rout = repl.obfuscate("SECRETUVABCDEFGH"); expect(rout).not.toContain("SECRETUV"); expect(rout).not.toContain("ABCDEFGH"); expect(repl.deobfuscate(rout)).toMatch(/ABCDEFGH$/); expect(repl.deobfuscate(rout)).not.toContain("SECRETUV"); expect(repl.obfuscate(rout)).toBe(rout); }); it("re-obfuscation is a fixed point when a greedy regex would spill into a trailing raw chunk around a prior-call placeholder", () => { // `[A-Z0-9]{8,12}` greedily matches `SECRETUVA` (9 chars) when the SECRETUV // placeholder is followed by the raw literal `A`. Previously the spillover // caused that short surrounding chunk to be minted into a fresh placeholder on // every subsequent obfuscate() call, drifting provider-visible history and the // prompt-cache prefix. The fix: a surrounding raw chunk is only obfuscated when // the placeholder's own deobfuscated value does NOT independently satisfy the // regex; when it does (greedy spillover), the surrounding bytes stay verbatim. const obf = new SecretObfuscator( [ { type: "plain", content: "ABCDEFGH" }, { type: "plain", content: "SECRETUV" }, { type: "regex", content: "[A-Z0-9]{8,12}" }, ], "Q".repeat(43), ); const first = obf.obfuscate("ZZZZZZZZABCDEFGHSECRETUVA"); // Core regression: re-obfuscation must be a fixed point. expect(obf.obfuscate(first)).toBe(first); // Round-trip must restore every byte of the original input. expect(obf.deobfuscate(first)).toBe("ZZZZZZZZABCDEFGHSECRETUVA"); // The trailing `A` must survive the first pass verbatim: SECRETUV alone // satisfies [A-Z0-9]{8,12}, so the short tail must not be placeholdered. expect(first.endsWith("A")).toBe(true); // Plain secrets must not appear in provider-visible output. expect(first).not.toContain("SECRETUV"); expect(first).not.toContain("ABCDEFGH"); }); it("first obfuscate() call already matches later calls when a bounded exact-quantifier regex spills past two adjacent placeholders", () => { // `[A-Z]{9}` (an EXACT quantifier, unlike the `{8,12}` RANGE case above) over // `ABCDEFGH` + `SECRETUV` + trailing raw `A`: the placeholder's OWN value // (`SECRETUV`, 8 chars) does NOT independently satisfy `{9}`, so this exercises // a different path than the independently-matching-spillover case. Previously // a cut-resolution resume point that landed exactly on the START of the // SECRETUV placeholder was handed straight to a fresh regex.exec instead of // being chained past it too: the FIRST obfuscate() call treated the still-raw // ABCDEFGH prefix as its own match and resumed right after it, leaving the // trailing `A` untouched — but a SECOND call, with ABCDEFGH already a // placeholder, started its match attempt inside the placeholder run, could not // be prefix-narrowed, and resumed mid-run instead of past it, exposing the // shorter tail `SECRETUV` + `A` to a clean match the first call never // attempted, minting a brand-new placeholder for `A` and drifting // provider-visible history / the prompt-cache prefix across the call-1-to-2 // transition. The fix chains the resume point through every immediately // adjacent generated-placeholder segment, so both calls land on the same // resolution from the very first pass. const obf = new SecretObfuscator( [ { type: "plain", content: "ABCDEFGH" }, { type: "plain", content: "SECRETUV" }, { type: "regex", content: "[A-Z]{9}" }, ], "Q".repeat(43), ); const first = obf.obfuscate("ZZZZZZZZABCDEFGHSECRETUVA"); // The trailing `A` must survive the FIRST pass verbatim. expect(first.endsWith("A")).toBe(true); // Core regression: re-obfuscation must already be a fixed point from the // very first call — this used to be false, turning `A` into a new placeholder // on the second call. expect(obf.obfuscate(first)).toBe(first); // Stable across multiple passes. expect(obf.obfuscate(obf.obfuscate(first))).toBe(first); // Round-trip must restore every byte of the original input. expect(obf.deobfuscate(first)).toBe("ZZZZZZZZABCDEFGHSECRETUVA"); }); it("keeps a default replace regex a fixed point on the first pass when raw text flanks a generated placeholder on both sides", () => { // `[A-Z]{9}` (exact quantifier) over raw `X` + plain secret `SECRETUV` // (minted fresh this call, 8 chars) + raw `Y`: the greedy left-to-right // match on the FIRST pass is exactly "XSECRETUV" (the 1-char raw prefix // plus the whole placeholder's expanded value), so the scan resumes right // past it and the trailing `Y` — only 1 char, short of the exact {9} // quantifier — was left raw and unredacted (`a#…#Y`). Once the prefix // redacts to a non-uppercase marker, a SECOND pass re-aligns the same // regex to match "SECRETUVY" instead (placeholder + the now-exposed raw // suffix) and redacts `Y` for the first time — drifting provider-visible // history and the prompt-cache prefix across the call-1-to-2 transition // (`a#…#Y` -> `a#…#a`). The fix must already redact both flanking chunks // on the very first pass so obfuscate() is a fixed point from the start. const obf = new SecretObfuscator( [ { type: "plain", content: "SECRETUV" }, { type: "regex", mode: "replace", content: "[A-Z]{9}" }, ], "Q".repeat(43), ); const first = obf.obfuscate("XSECRETUVY"); // The plain secret must never survive in provider-visible output. expect(first).not.toContain("SECRETUV"); // Core regression: the trailing raw suffix must already be redacted on // the FIRST pass instead of surviving verbatim until a second call // re-aligns the match around it. expect(first.endsWith("Y")).toBe(false); // Core regression: re-obfuscation must already be a fixed point from the // very first call — this used to be false (`a#…#Y` -> `a#…#a`). expect(obf.obfuscate(first)).toBe(first); // Stable across multiple passes. expect(obf.obfuscate(obf.obfuscate(first))).toBe(first); // Deobfuscating restores the plain secret (reversible), but the regex's // one-way replace-mode redaction never restores the raw flanking bytes // it consumed — that is the correct, appropriate behavior for replace // mode. const restored = obf.deobfuscate(first); expect(restored).toContain("SECRETUV"); expect(restored).not.toBe("XSECRETUVY"); }); it("redacts a two-sided independently matching chunk instead of leaking it as spillover", () => { // `\b[A-Z]{8}\b|[A-Z]{17}` union regex, prior-call placeholder for `SECRETUV` // flanked by prefix `ABCDEFGH` (independently matches `\b[A-Z]{8}\b`) and // suffix `I` (matches nothing alone). The straddling match is the 17-char // run `ABCDEFGH` + `SECRETUV` + `I`, which previously tested the concatenated // outside chunks `"ABCDEFGH" + "I"` = `"ABCDEFGHI"` against the regex — that // concatenation erases the placeholder-token boundary between the chunks, so // neither alternative matches (9 chars, and `H` is no longer at a word // boundary), and the whole match was left verbatim, leaking `ABCDEFGH` // unredacted. The fix tests each outside chunk in its own real, flanked // context: `ABCDEFGH` alone still independently matches `\b[A-Z]{8}\b`, so it // is redacted instead of leaked. const obf = new SecretObfuscator( [ { type: "plain", content: "SECRETUV" }, { type: "regex", content: "\\b[A-Z]{8}\\b|[A-Z]{17}" }, ], "Q".repeat(43), ); const placeholdered = obf.obfuscate("SECRETUV"); const second = obf.obfuscate(`ABCDEFGH${placeholdered}I`); // Core regression: the independently matching prefix must not survive // verbatim in provider-visible output. expect(second).not.toContain("ABCDEFGH"); expect(second).not.toContain("SECRETUV"); // Every byte must still round-trip. expect(obf.deobfuscate(second)).toBe("ABCDEFGHSECRETUVI"); // Redacting the prefix must itself be a fixed point. expect(obf.obfuscate(second)).toBe(second); }); it("leaves a genuine one-sided spillover chunk verbatim under the same union regex", () => { // Contrast case for the fix above: the trailing chunk `IJKLMNOPQ` (9 chars) // does NOT independently match `\b[A-Z]{8}\b|[A-Z]{17}` on its own, and only // forms a match by bridging across the `SECRETUV` placeholder (`SECRETUV` + // `IJKLMNOPQ` = 17 chars). This is genuine greedy spillover, not independent // secret-shaped content, so per-chunk testing must still leave it verbatim // and a fixed point — the fix must not turn spillover into a false positive. const obf = new SecretObfuscator( [ { type: "plain", content: "SECRETUV" }, { type: "regex", content: "\\b[A-Z]{8}\\b|[A-Z]{17}" }, ], "Q".repeat(43), ); const placeholdered = obf.obfuscate("SECRETUV"); const second = obf.obfuscate(`${placeholdered}IJKLMNOPQ`); expect(second.endsWith("IJKLMNOPQ")).toBe(true); expect(obf.deobfuscate(second)).toBe("SECRETUVIJKLMNOPQ"); expect(obf.obfuscate(second)).toBe(second); }); it("redacts a lookbehind-anchored prefix tested in its real source context instead of leaking it", () => { // `(?<=api=)[0-9]{8}[A-Z]{8}|(?<=api=)[0-9]{8}|[A-Z]{8}` combined with a // prior-call placeholder for `ABCDEFGH` (independently matches the bare // `[A-Z]{8}` alternative). Re-obfuscating `api=12345678` + placeholder makes // the full match (placeholder expanded) `12345678ABCDEFGH`, preceded by the // literal `api=`. The outside chunk is `12345678`. Testing that chunk as an // ISOLATED slice (starting a fresh string at index 0) loses the real `api=` // bytes immediately before it in the source, so `(?<=api=)[0-9]{8}` never // fires, none of the three alternatives match a bare "12345678", and the // digits were wrongly left verbatim — a lookbehind-blind leak. The fix tests // the chunk against the regex at its REAL position in the full text, where // the lookbehind correctly sees the preceding `api=` and the chunk is // identified as an independent match, so it is redacted like any other // secret-shaped content. const obf = new SecretObfuscator( [ { type: "plain", content: "ABCDEFGH" }, { type: "regex", content: "(?<=api=)[0-9]{8}[A-Z]{8}|(?<=api=)[0-9]{8}|[A-Z]{8}" }, ], "Q".repeat(43), ); const placeholdered = obf.obfuscate("ABCDEFGH"); const second = obf.obfuscate(`api=12345678${placeholdered}`); // Core regression: the lookbehind-qualified digits must not survive // verbatim in provider-visible output. expect(second).not.toContain("12345678"); expect(second).not.toContain("ABCDEFGH"); // Every byte must still round-trip. expect(obf.deobfuscate(second)).toBe("api=12345678ABCDEFGH"); // Redacting the prefix must itself be a fixed point. expect(obf.obfuscate(second)).toBe(second); }); it("redacts a suffix whose lookbehind only resolves against the expanded scan context, not the literal placeholder token", () => { // `ABCDEFGHSECRET|(?<=ABCDEFGH)SECRET|ABCDEFGH` combined with a prior-call // placeholder for `ABCDEFGH` (independently matches the bare `ABCDEFGH` // alternative). Re-obfuscating placeholder + `SECRET` makes the full match // (placeholder expanded) `ABCDEFGHSECRET`; the outside chunk is `SECRET`. // The ONLY way `SECRET` independently matches on its own is via the second // alternative's lookbehind `(?<=ABCDEFGH)SECRET` — but the bytes immediately // preceding it in the actual text are the LITERAL placeholder token (`#…#`), // not `ABCDEFGH`, so testing the chunk in its literal-token context alone // never satisfies the lookbehind and wrongly reports no independent match // (unlike the two-sided-chunk test above, which literal context alone does // catch). Testing the same chunk against the EXPANDED scan context — where // the placeholder is resolved back to `ABCDEFGH` — lets the lookbehind see // its real preceding bytes and correctly reports an independent match. // Without that expanded-context check unioned with the literal-context one, // `SECRET` was wrongly treated as spillover and leaked verbatim beside the // placeholder. const obf = new SecretObfuscator( [ { type: "plain", content: "ABCDEFGH" }, { type: "regex", content: "ABCDEFGHSECRET|(?<=ABCDEFGH)SECRET|ABCDEFGH" }, ], "Q".repeat(43), ); const placeholdered = obf.obfuscate("ABCDEFGH"); const second = obf.obfuscate(`${placeholdered}SECRET`); // Core regression: the lookbehind-qualified suffix must not survive // verbatim in provider-visible output. expect(second).not.toContain("SECRET"); expect(second).not.toContain("ABCDEFGH"); // Every byte must still round-trip. expect(obf.deobfuscate(second)).toBe("ABCDEFGHSECRET"); // Redacting the suffix must itself be a fixed point. expect(obf.obfuscate(second)).toBe(second); }); it("redacts a cut prefix using placeholder right context instead of leaking it", () => { const obf = new SecretObfuscator( [ { type: "plain", content: "ABCDEFGH" }, { type: "regex", content: "SECRETUVABCD|SECRETUV(?=ABCD)|[A-Z]{8}" }, ], "Q".repeat(43), ); const placeholdered = obf.obfuscate("ABCDEFGH"); const second = obf.obfuscate(`SECRETUV${placeholdered}`); // The prefix match depends on lookahead supplied by the expanded placeholder. // It must still be redacted while the placeholder's own bytes stay atomic. expect(second).not.toContain("SECRETUV"); expect(second).toContain(placeholdered); expect(obf.deobfuscate(second)).toBe("SECRETUVABCDEFGH"); expect(obf.obfuscate(second)).toBe(second); }); it("redacts a cut suffix using placeholder left context instead of leaking it", () => { const key = "AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA1n"; const entries = [ { type: "plain" as const, content: "ABCDEFGH" }, { type: "regex" as const, mode: "replace" as const, content: "(?<=ABCD)[A-Z]{8}" }, ]; const obf = new SecretObfuscator(entries, key); const placeholdered = obf.obfuscate("ABCDEFGH"); const second = obf.obfuscate("ABCDEFGHIJKL"); // The suffix match depends on lookbehind supplied by the expanded placeholder. // It must still be redacted while the placeholder's own bytes stay atomic. expect(second).not.toContain("IJKL"); expect(second).toContain(placeholdered); expect(obf.obfuscate(second)).toBe(second); const restarted = new SecretObfuscator(entries, key); expect(restarted.obfuscate(second)).toBe(second); }); it("obfuscates a short cut suffix covered by a full-length placeholder-context match", () => { const obf = new SecretObfuscator( [ { type: "plain", content: "ABCDEFGH" }, { type: "regex", content: "(?<=ABCD)[A-Z]{8}" }, ], "Q".repeat(43), ); const second = obf.obfuscate("ABCDEFGHIJKL"); expect(second).not.toContain("IJKL"); expect(obf.deobfuscate(second)).toBe("ABCDEFGHIJKL"); expect(obf.obfuscate(second)).toBe(second); }); it("keeps default replace markers stable when a lookbehind match spills into a prior placeholder", () => { const key = "AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA1n"; const entries = [ { type: "plain" as const, content: "ABCDEFGH" }, { type: "regex" as const, mode: "replace" as const, content: "(?<=api=)[0-9]{8}[A-Z]{8}|(?<=api=)[0-9]{8}|[A-Z]{8}", }, ]; const obf = new SecretObfuscator(entries, key); const placeholdered = obf.obfuscate("ABCDEFGH"); const persisted = obf.obfuscate(`api=12345678${placeholdered}`); expect(persisted).not.toContain("12345678"); expect(persisted).not.toContain("ABCDEFGH"); expect(persisted).toContain(placeholdered); expect(obf.obfuscate(persisted)).toBe(persisted); const restarted = new SecretObfuscator(entries, key); expect(restarted.obfuscate(persisted)).toBe(persisted); }); it("preserves a same-call fresh placeholder's origin across sequential replace-mode regex entries", () => { // Codex review on PR #2735 ("Preserve fresh placeholder origin after regex // redactions"): `ABCDEFGH` (obfuscate mode) + `[A-Z]{8}` (replace) + // `[A-Z]{8,14}` (replace), all discovered in a SINGLE obfuscate() call on // `ABCDEFGHSECRET`. Step 2 turns `ABCDEFGH` into a placeholder tagged "F" // (fresh this call). The first regex entry's redaction pass matches that // placeholder token exactly and preserves it untouched — but blanket-tagging // the whole redacted span "I" mislabeled the preserved placeholder as // prior-call. That false "I" tag then made the SECOND regex entry treat the // placeholder as eligible for the greedy-spillover fixed-point skip: its // deobfuscated inner value (`ABCDEFGH`, 8 chars) independently matches // `[A-Z]{8,14}` while the outside chunk (`SECRET`, 6 chars) does not, so // redaction was skipped entirely, leaking `SECRET` verbatim in the output. // The fix threads the origin tag through the redaction helpers so a // preserved placeholder keeps its own "F" tag, and the second entry // correctly redacts `SECRET` instead of treating it as spillover. const obf = new SecretObfuscator( [ { type: "plain", content: "ABCDEFGH" }, { type: "regex", content: "[A-Z]{8}", mode: "replace" }, { type: "regex", content: "[A-Z]{8,14}", mode: "replace" }, ], "Q".repeat(43), ); const result = obf.obfuscate("ABCDEFGHSECRET"); // Core regression: the trailing secret must not survive verbatim in // provider-visible output. expect(result).not.toContain("SECRET"); // Re-obfuscating the already-redacted output must be a fixed point. expect(obf.obfuscate(result)).toBe(result); }); it("does not preserve a raw outside chunk merely because its bytes match a generated replacement minted elsewhere", () => { // PR #4636 review: default replace-mode preservation checked ONLY // `#generatedReplaceChunks.has(match.inputPlaceholderOutside)` — a set of // every replacement value this obfuscator has EVER emitted, keyed by VALUE // across its whole lifetime, not scoped to the current call. A raw outside // chunk that merely happens to share bytes with a replacement minted for a // completely different secret elsewhere was treated as "already redacted" // and skipped, even though this exact span was never touched by this // obfuscator. The fix additionally requires the outside chunk's own origin // to carry an "F" tag (generated THIS call) before trusting the value // collision, so raw input can never be waved through just because some // unrelated redaction happened to produce the same bytes. const obf = new SecretObfuscator( [ { type: "plain", content: "ABCDEFGH" }, { type: "regex", mode: "replace", content: "[A-Z]{9}" }, { type: "regex", mode: "replace", content: "X" }, ], "Q".repeat(43), ); const placeholder = obf.obfuscate("ABCDEFGH"); // Seed `#generatedReplaceChunks` with "Z" via an UNRELATED occasion: a raw // "X" directly abutting the placeholder forms "XABCDEFGH" (9 chars), so the // 1-char outside chunk is redacted through the length-keyed chunk sentinel // (`#generateReplacement` always probes "Z" first for a 1-char remainder and // records it in `#generatedReplaceChunks` even when a distinct non-matching // value is ultimately emitted for THIS chunk). obf.obfuscate(`X${placeholder}`); // A brand-new, never-before-seen raw "Z" now trails the SAME placeholder: // expanded, `ABCDEFGH` + `Z` forms `ABCDEFGHZ` — another 9-char match. This // "Z" was never generated by this obfuscator; it only coincidentally equals // a value already sitting in `#generatedReplaceChunks`. const result = obf.obfuscate(`${placeholder}Z`); // Core regression: the raw trailing `Z` must be redacted on this FIRST // pass, not preserved verbatim just because "Z" is already a member of // `#generatedReplaceChunks` from an unrelated redaction elsewhere. expect(result).not.toBe(`${placeholder}Z`); expect(result.endsWith("Z")).toBe(false); expect(result.startsWith(placeholder)).toBe(true); // Must already be a fixed point on the very first pass. expect(obf.obfuscate(result)).toBe(result); // The reversible obfuscate-mode placeholder still recovers ABCDEFGH; the // one-way replace-mode marker consumed for the raw `Z` is never restored. const restored = obf.deobfuscate(result); expect(restored).toContain("ABCDEFGH"); expect(restored).not.toBe("ABCDEFGHZ"); }); it("never mints a placeholder equal to another configured secret's literal value", () => { // Codex review on PR #2735 ("Reject placeholders that equal configured // secret values"): `#placeholderConflicts` only checked the internal // deobfuscation map, not the full set of configured plain-secret literals. // Discover the exact keyed placeholder `ABCDEFGH` mints under a fixed key, // then configure THAT placeholder string as a second, unrelated plain // secret ("B"). Its own plain-secret redaction pass (sorted by length, run // once per obfuscate() call) already completed before `ABCDEFGH`'s // placeholder existed in the text, so a naive conflict check never catches // the collision — the newly-minted placeholder becomes a verbatim, // provider-visible copy of B's raw secret. Covers both entry orderings // since the bug depended on which secret's redaction pass ran first. const key = "Q".repeat(43); const discoveryObf = new SecretObfuscator([{ type: "plain", content: "ABCDEFGH" }], key); const collidingPlaceholder = discoveryObf.obfuscate("ABCDEFGH"); expect(collidingPlaceholder).toMatch(/^\$\$[A-Z0-9]+(?::[ULCM])?\$\$$/); for (const entries of [ [ { type: "plain" as const, content: "ABCDEFGH" }, { type: "plain" as const, content: collidingPlaceholder }, ], [ { type: "plain" as const, content: collidingPlaceholder }, { type: "plain" as const, content: "ABCDEFGH" }, ], ]) { const obf = new SecretObfuscator(entries, key); const input = "value=ABCDEFGH other=stuff"; const out = obf.obfuscate(input); // Core regression: B's raw literal must never survive verbatim, // disguised as A's placeholder, in provider-visible output. expect(out).not.toContain(collidingPlaceholder); // Deobfuscation still restores the original input. expect(obf.deobfuscate(out)).toBe(input); // Re-obfuscating the already-redacted output must be a fixed point. expect(obf.obfuscate(out)).toBe(out); } }); it("keeps regex placeholders stable when inner friendly names change", () => { const sharedKey = "E".repeat(43); const before = new SecretObfuscator( [ { type: "plain", content: "abcdefgh", friendlyName: "old" }, { type: "regex", content: "api_key=\\S+", friendlyName: "api-key" }, ], sharedKey, ); const persisted = before.obfuscate("api_key=abcdefghXYZ"); const after = new SecretObfuscator( [ { type: "plain", content: "abcdefgh", friendlyName: "new" }, { type: "regex", content: "api_key=\\S+", friendlyName: "api-key" }, ], sharedKey, ); expect(after.obfuscate("api_key=abcdefghXYZ")).toBe(persisted); expect(after.deobfuscate(persisted)).toBe("api_key=abcdefghXYZ"); }); it("keeps default replace markers stable when friendly placeholder aliases advance scan context", () => { const sharedKey = "E".repeat(43); const regex = "api_key=[a-z]{8}[A-Za-z0-9]{3}|(?<=abcdefgh)[A-Za-z0-9]{3}"; const before = new SecretObfuscator( [ { type: "plain", content: "abcdefgh", friendlyName: "old" }, { type: "regex", mode: "replace", content: regex }, ], sharedKey, ); const oldPlaceholder = before.obfuscate("abcdefgh"); const after = new SecretObfuscator( [ { type: "plain", content: "abcdefgh", friendlyName: "new" }, { type: "regex", mode: "replace", content: regex }, ], sharedKey, ); const persisted = after.obfuscate(`api_key=${oldPlaceholder}XYZ`); expect(persisted).toContain(oldPlaceholder); expect(persisted).not.toContain("XYZ"); expect(after.obfuscate(persisted)).toBe(persisted); }); it("redacts literal hash-delimited text inside regex matches", () => { const obfuscator = new SecretObfuscator( [{ type: "regex", content: "api_key=\\S+", friendlyName: "api-key" }], "F".repeat(43), ); const obfuscated = obfuscator.obfuscate("api_key=#XRRS#"); expect(obfuscated).toMatch(/^\$\$APIKEY_[A-Z0-9]+(?::[ULCM])?\$\$$/); expect(obfuscator.deobfuscate(obfuscated)).toBe("api_key=#XRRS#"); }); it("redacts replace-mode regex spans around generated placeholders", () => { const obfuscator = new SecretObfuscator( [ { type: "plain", content: "SECRETUV" }, { type: "regex", mode: "replace", content: "[A-Z0-9]{8,}", replacement: "REDACTED" }, ], "A".repeat(43), ); const obfuscated = obfuscator.obfuscate("SECRETUVX1"); expect(obfuscated).toMatch(/^\$\$[A-Z0-9]+:U\$\$REDACTED$/); expect(obfuscated).not.toMatch(/X1$/); expect(obfuscator.deobfuscate(obfuscated)).toBe("SECRETUVREDACTED"); }); it("redacts bounded replace-mode regex suffixes after generated placeholders", () => { const obfuscator = new SecretObfuscator( [ { type: "plain", content: "SECRETUV" }, { type: "regex", mode: "replace", content: "[A-Z0-9]{10}", replacement: "REDACTED" }, ], "B".repeat(43), ); const obfuscated = obfuscator.obfuscate("SECRETUVX1"); // The 8-char SECRETUVX1 redacts to one placeholder + REDACTED; assert the `X1` // suffix is gone via end-anchored structure, not substring absence — the // random keyed base can itself contain the two chars "X1". expect(obfuscated).toMatch(/^\$\$[A-Z0-9]+:U\$\$REDACTED$/); expect(obfuscated).not.toMatch(/X1$/); expect(obfuscator.deobfuscate(obfuscated)).toBe("SECRETUVREDACTED"); }); it("emits a custom replacement once around a generated placeholder", () => { const obfuscator = new SecretObfuscator( [ { type: "plain", content: "abcdefgh" }, { type: "regex", mode: "replace", content: "api_key=\\S+", replacement: "REDACTED" }, ], "C".repeat(43), ); const obfuscated = obfuscator.obfuscate("api_key=abcdefghXYZ"); // A custom replacement is a single redaction marker for the whole match, so // it must not be duplicated on both sides of the preserved placeholder (the // bug produced `REDACTED#…#REDACTED`). Asserted by structure plus an // end-anchored guard rather than a base-collidable substring count. expect(obfuscated).toMatch(/^REDACTED\$\$[A-Z0-9]+:L\$\$$/); expect(obfuscated).not.toMatch(/REDACTED$/); expect(obfuscated).not.toContain("api_key="); expect(obfuscator.deobfuscate(obfuscated)).toBe("REDACTEDabcdefgh"); }); it("is idempotent when re-obfuscating already-obfuscated text", () => { // The SDK obfuscates messages in both convertToLlm and transformProviderContext, // and prior-turn messages re-enter every turn, so obfuscate() must be a fixed // point. Re-running it on its own output must not re-redact around an existing // placeholder (regression: `#…#REDACTED` -> `#…#REDACTEDDACTED`). const replace = new SecretObfuscator( [ { type: "plain", content: "SECRETUV" }, { type: "regex", mode: "replace", content: "[A-Z0-9]{10}", replacement: "REDACTED" }, ], "D".repeat(43), ); const replaceOnce = replace.obfuscate("SECRETUVX1"); expect(replaceOnce).toMatch(/^\$\$[A-Z0-9]+:U\$\$REDACTED$/); expect(replace.obfuscate(replaceOnce)).toBe(replaceOnce); expect(replace.obfuscate(replace.obfuscate(replaceOnce))).toBe(replaceOnce); expect(replace.deobfuscate(replaceOnce)).toBe("SECRETUVREDACTED"); // Custom replacement spanning a placeholder must also stay a fixed point. const custom = new SecretObfuscator( [ { type: "plain", content: "abcdefgh" }, { type: "regex", mode: "replace", content: "api_key=\\S+", replacement: "REDACTED" }, ], "E".repeat(43), ); const customOnce = custom.obfuscate("api_key=abcdefghXYZ"); expect(custom.obfuscate(customOnce)).toBe(customOnce); expect(custom.deobfuscate(customOnce)).toBe("REDACTEDabcdefgh"); // Obfuscate-mode regex spanning a placeholder is a fixed point too. const obf = new SecretObfuscator( [ { type: "plain", content: "abcdefgh" }, { type: "regex", content: "api_key=[A-Za-z0-9]{11}", friendlyName: "api-key" }, ], "F".repeat(43), ); const obfOnce = obf.obfuscate("api_key=abcdefghXYZ"); expect(obf.obfuscate(obfOnce)).toBe(obfOnce); expect(obf.deobfuscate(obfOnce)).toBe("api_key=abcdefghXYZ"); }); it("cross-matches a fresh placeholder whose token already appears in the input", () => { const obf = new SecretObfuscator( [ { type: "plain", content: "abcdefgh" }, { type: "regex", mode: "replace", content: "api_key=\\S+", replacement: "REDACTED" }, ], "G".repeat(43), ); const token = obf.obfuscate("abcdefgh"); expect(token).toMatch(/^\$\$[A-Z0-9]+:L\$\$$/); // Input carries the prior token literally AND a fresh api_key=abcdefghXYZ (raw `abc`). // The fresh occurrence must still be redacted (XYZ gone) while the prior token is // preserved; range-based origin tracking distinguishes the two same-token spans, // where a token-value guard would skip both and leak XYZ. const out = obf.obfuscate(`${token} api_key=abcdefghXYZ`); expect(out).toBe(`${token} REDACTED${token}`); expect(obf.deobfuscate(out)).toBe("abcdefgh REDACTEDabcdefgh"); expect(obf.obfuscate(out)).toBe(out); // still a fixed point }); it("redacts new surrounding bytes around a prior-call placeholder without re-redacting markers", () => { const obf = new SecretObfuscator( [ { type: "plain", content: "abcdefgh" }, { type: "regex", mode: "replace", content: "api_key=\\S+", replacement: "REDACTED" }, ], "H".repeat(43), ); // Simulate a session where `abc` was obfuscated in an earlier turn (the token // is a prior-call/input placeholder) and the regex now re-enters text where // `api_key=` + raw `XYZ` still surround that token. The prior placeholder is // preserved, but the genuinely-new surrounding bytes (`api_key=`, `XYZ`) must // be redacted — not dropped, which would leak `XYZ` to the provider. const token = obf.obfuscate("abcdefgh"); const out = obf.obfuscate(`api_key=${token}XYZ`); expect(out).toBe(`REDACTED${token}`); expect(out).not.toContain("XYZ"); expect(out).not.toContain("api_key="); expect(obf.deobfuscate(out)).toBe("REDACTEDabcdefgh"); // Re-obfuscating the redacted output is a fixed point: the marker `REDACTED` // does not independently satisfy `api_key=\S+`, so nothing grows. expect(obf.obfuscate(out)).toBe(out); }); it("redacts bounded replace regex remainders around prior placeholders", () => { const obf = new SecretObfuscator( [ { type: "plain", content: "abcdefgh" }, { type: "regex", mode: "replace", content: "api_key=[A-Za-z0-9]{11}", replacement: "REDACTED" }, ], "I".repeat(43), ); const token = obf.obfuscate("abcdefgh"); const out = obf.obfuscate(`api_key=${token}XYZ`); expect(out).toBe(`REDACTED${token}`); expect(out).not.toContain("XYZ"); expect(out).not.toContain("api_key="); expect(obf.deobfuscate(out)).toBe("REDACTEDabcdefgh"); expect(obf.obfuscate(out)).toBe(out); }); it("redacts custom replacement prefixes that are raw regex remainders", () => { const obf = new SecretObfuscator( [ { type: "plain", content: "SECRETUV" }, { type: "regex", mode: "replace", content: "[A-Z0-9]{9}", replacement: "REDACTED" }, ], "K".repeat(43), ); const token = obf.obfuscate("SECRETUV"); const out = obf.obfuscate(`${token}R`); expect(out).toBe(`${token}REDACTED`); expect(obf.deobfuscate(out)).toBe("SECRETUVREDACTED"); expect(obf.obfuscate(out)).toBe(out); }); it("redacts trailing bytes after existing custom replacement markers", () => { const obf = new SecretObfuscator( [ { type: "plain", content: "SECRETUV" }, { type: "regex", mode: "replace", content: "[A-Z0-9]{17}", replacement: "REDACTED" }, ], "L".repeat(43), ); const token = obf.obfuscate("SECRETUV"); const out = obf.obfuscate(`${token}REDACTEDX`); expect(out).toBe(`${token}REDACTED`); expect(obf.obfuscate(out)).toBe(out); }); it("keeps bounded custom replacement matches idempotent around preserved placeholders", () => { const obf = new SecretObfuscator( [ { type: "plain", content: "SECRETUV" }, { type: "regex", mode: "replace", content: "[A-Z0-9]{8,12}", replacement: "REDACTED" }, ], "M".repeat(43), ); const out = obf.obfuscate("XSECRETUVREDACTED"); expect(out).toMatch(/^REDACTED\$\$[A-Z0-9]+:U\$\$REDACTED$/); expect(obf.obfuscate(out)).toBe(out); expect(obf.deobfuscate(out)).toBe("REDACTEDSECRETUVREDACTED"); }); it("redacts default replace regex remainders around prior placeholders", () => { const obf = new SecretObfuscator( [ { type: "plain", content: "abcdefgh" }, { type: "regex", mode: "replace", content: "api_key=[A-Za-z0-9]{11}" }, ], "J".repeat(43), ); const token = obf.obfuscate("abcdefgh"); const out = obf.obfuscate(`api_key=${token}XYZ`); expect(out).toContain(token); expect(out).not.toContain("XYZ"); expect(out).not.toContain("api_key="); expect(obf.obfuscate(out)).toBe(out); }); it("redacts bounded regex remainders around renamed friendly placeholders", () => { const sharedKey = "J".repeat(43); const before = new SecretObfuscator([{ type: "plain", content: "abcdefgh", friendlyName: "old" }], sharedKey); const oldToken = before.obfuscate("abcdefgh"); const after = new SecretObfuscator( [ { type: "plain", content: "abcdefgh", friendlyName: "new" }, { type: "regex", mode: "replace", content: "api_key=[A-Za-z0-9]{11}" }, ], sharedKey, ); const out = after.obfuscate(`api_key=${oldToken}XYZ`); expect(out).toContain(oldToken); expect(out).not.toContain("api_key="); expect(out).not.toContain("XYZ"); expect(after.obfuscate(out)).toBe(out); }); it("redacts default replace raw suffixes after prior placeholders", () => { const obf = new SecretObfuscator( [ { type: "plain", content: "SECRETUV" }, { type: "regex", mode: "replace", content: "[A-Z0-9]{10}" }, ], "N".repeat(43), ); const token = obf.obfuscate("SECRETUV"); const out = obf.obfuscate(`${token}X1`); expect(out).toContain(token); expect(out).not.toContain("X1"); expect(obf.obfuscate(out)).toBe(out); }); it("redacts raw default replace values that look like generated sentinels", () => { const obf = new SecretObfuscator([{ type: "plain", content: "ZZTOPSECRET", mode: "replace" }], "Q".repeat(43)); const out = obf.obfuscate("ZZTOPSECRET"); expect(out).not.toBe("ZZTOPSECRET"); expect(out).toHaveLength("ZZTOPSECRET".length); }); it("does not emit a plain replace secret equal to the Z/ZZ sentinel unchanged", () => { for (const content of ["Z", "ZZ"]) { const obf = new SecretObfuscator([{ type: "plain", content, mode: "replace" }], "Q".repeat(43)); const out = obf.obfuscate(content); expect(out).not.toBe(content); expect(out).toHaveLength(content.length); // Replace mode is one-way; re-obfuscating the emitted value is a fixed point. expect(obf.obfuscate(out)).toBe(out); } }); it("does not emit a default replace regex match equal to the Z/ZZ sentinel unchanged", () => { // A literal regex whose match is exactly the sentinel: the perturbed value is // not re-matched, so it is emitted distinct AND stays a fixed point. for (const content of ["Z", "ZZ"]) { const obf = new SecretObfuscator([{ type: "regex", mode: "replace", content }], "Q".repeat(43)); const out = obf.obfuscate(content); expect(out).not.toBe(content); expect(out).toHaveLength(content.length); expect(obf.obfuscate(out)).toBe(out); } }); it("redacts a self-matching sentinel regex to a stable nonmatching value", () => { // A regex that also matches the single A/B perturbation still has same-length // values it does NOT match (a lowercase pair for [A-Z]{2}, an A/Z-free pair for // Z+). The bounded search finds one, so the sentinel is redacted to a value the // regex never re-matches: leak-free AND a fixed point under re-obfuscation. for (const content of ["Z+", "[A-Z]{2}"]) { const obf = new SecretObfuscator([{ type: "regex", mode: "replace", content }], "Q".repeat(43)); const out = obf.obfuscate("ZZ"); expect(out).not.toBe("ZZ"); expect(out).toHaveLength(2); expect(obf.obfuscate(out)).toBe(out); expect(obf.obfuscate(obf.obfuscate(out))).toBe(out); } }); it("searches past the first perturbation when it also matches the regex", () => { // Regression for a regex that matches both the sentinel and its single A/B // perturbation: `Z|A`/`[AZ]` match `Z` and `A`, so the old guard kept the raw // `Z` and shipped it. A nonmatching same-length value (`B`) must be found. for (const content of ["Z|A", "[AZ]"]) { const obf = new SecretObfuscator([{ type: "regex", mode: "replace", content }], "Q".repeat(43)); const out = obf.obfuscate("Z"); expect(out).not.toBe("Z"); expect(out).toHaveLength(1); expect(obf.obfuscate(out)).toBe(out); } }); it("falls back to punctuation when a regex matches every alphanumeric sentinel candidate", () => { const obf = new SecretObfuscator([{ type: "regex", mode: "replace", content: "[A-Za-z0-9]{2}" }], "Q".repeat(43)); const out = obf.obfuscate("ZZ"); expect(out).not.toBe("ZZ"); expect(out).toHaveLength(2); expect(/^[A-Za-z0-9]{2}$/.test(out)).toBe(false); expect(obf.obfuscate(out)).toBe(out); }); it("exhausts two-character fallback candidates before keeping the sentinel", () => { const obf = new SecretObfuscator([{ type: "regex", mode: "replace", content: "[A-Za-z0-9]." }], "Q".repeat(43)); const out = obf.obfuscate("ZZ"); expect(out).not.toBe("ZZ"); expect(out).toHaveLength(2); expect(/[A-Za-z0-9]./.test(out)).toBe(false); expect(obf.obfuscate(out)).toBe(out); }); it("samples every leading character class before giving up on three-character collisions", () => { const obf = new SecretObfuscator( [{ type: "regex", mode: "replace", content: "[A-Za-z0-9].{2}" }], "Q".repeat(43), ); const out = obf.obfuscate("ZZc"); expect(out).not.toBe("ZZc"); expect(out).toHaveLength(3); expect(/[A-Za-z0-9].{2}/.test(out)).toBe(false); expect(obf.obfuscate(out)).toBe(out); }); it("exhausts three-character fallback candidates when the nonmatching byte must be last", () => { const obf = new SecretObfuscator( [{ type: "regex", mode: "replace", content: ".{2}[A-Za-z0-9]" }], "Q".repeat(43), ); const out = obf.obfuscate("ZZc"); expect(out).not.toBe("ZZc"); expect(out).toHaveLength(3); expect(/.{2}[A-Za-z0-9]/.test(out)).toBe(false); expect(obf.obfuscate(out)).toBe(out); }); it("exhausts longer fallback candidates when the nonmatching byte must be last", () => { const obf = new SecretObfuscator( [{ type: "regex", mode: "replace", content: ".{4}[A-Za-z0-9]" }], "Q".repeat(43), ); const out = obf.obfuscate("ZZLB6"); expect(out).not.toBe("ZZLB6"); expect(out).toHaveLength(5); expect(/.{4}[A-Za-z0-9]/.test(out)).toBe(false); expect(obf.obfuscate(out)).toBe(out); }); it("tries multi-character fallback replacements when one changed byte is still matchable", () => { const obf = new SecretObfuscator([{ type: "regex", mode: "replace", content: "[A-Za-z0-9].*" }], "Q".repeat(43)); const out = obf.obfuscate("ZZLB6"); expect(out).not.toBe("ZZLB6"); expect(out).toHaveLength(5); expect(/[A-Za-z0-9].*/.test(out)).toBe(false); expect(obf.obfuscate(out)).toBe(out); }); it("redacts a default replace regex that matches every non-whitespace candidate", () => { // `\S{5}` matches every non-whitespace value, so the alphanumeric and // punctuation candidates are all exhausted. A same-length whitespace run is a // stable nonmatching redaction, so the colliding sentinel value is replaced // rather than shipped raw to the provider. const obf = new SecretObfuscator([{ type: "regex", mode: "replace", content: "\\S{5}" }], "Q".repeat(43)); const out = obf.obfuscate("ZZLB6"); expect(out).not.toBe("ZZLB6"); expect(out).toHaveLength(5); expect(/\S{5}/.test(out)).toBe(false); expect(obf.obfuscate(out)).toBe(out); }); it("redacts a replace regex that also matches all-space and all-tab runs via a mixed marker", () => { // `(?:\S{5}| {5}|\t{5})` matches every non-whitespace run AND a full space or // tab run, so neither the punctuation candidates nor a full whitespace run is a // stable redaction. A mixed marker (one whitespace byte among filler, e.g. // ` AAAA`) breaks every fixed-length run, so the colliding sentinel value is // still redacted to a stable nonmatching value instead of shipped raw. const obf = new SecretObfuscator( [{ type: "regex", mode: "replace", content: "(?:\\S{5}| {5}|\\t{5})" }], "Q".repeat(43), ); const out = obf.obfuscate("ZZLB6"); expect(out).not.toBe("ZZLB6"); expect(out).toHaveLength(5); expect(/(?:\S{5}| {5}|\t{5})/.test(out)).toBe(false); expect(obf.obfuscate(out)).toBe(out); }); it("drops a replace regex entirely when it can never redact a 1-2 char match distinctly from itself", () => { // A match-everything regex has no nonmatching same-length redaction for a // 1-2 char value, so `findNonMatchingReplacement` would exhaust even against // an isolated probe drawn from the fallback's own alphabet — the regex has // already proven it matches literally every candidate that alphabet could // produce. Accepting the entry would mean the only "redaction" available is // the raw match returned unchanged, which is worse than not registering the // entry at all. The constructor now drops such entries silently at // construction (`regexHasUnresolvableShortMatchFallback`), so with no other // configured secret, `hasSecrets()` is false and the input passes through // completely unobfuscated. for (const content of [".", "[\\s\\S]", "[\\s\\S]{2}"]) { const obf = new SecretObfuscator([{ type: "regex", mode: "replace", content }], "Q".repeat(43)); const input = content === "[\\s\\S]{2}" ? "ZZ" : "Z"; expect(obf.hasSecrets()).toBe(false); expect(obf.obfuscate(input)).toBe(input); } }); it("resolves a three-character match-everything regex without the removed exhaustive sweep", () => { // Regression: `findNonMatchingReplacement` used to special-case `len <= 3` // with a fully exhaustive search over every `base**length` candidate // (base = 90, so 90**3 = 729000) before falling back to the whitespace/keyed // marker. A 3-byte default replace regex that matches every candidate // (`[\s\S]{3}`) rejects the ENTIRE sweep on every single match, so this exact // shape burned the whole 729000-candidate search (measured ~70ms on this // machine) on one redaction. The fix drops the special case so length <= 3 // goes through the same bounded single-position-substitution search already // used for longer values (O(length * 90)), measured at ~1ms for this case — // correctness is unchanged (the caller's keyed marker is still the fixed // point once the bounded search exhausts), but the pathological cost is gone. const obf = new SecretObfuscator([{ type: "regex", mode: "replace", content: "[\\s\\S]{3}" }], "Q".repeat(43)); const start = performance.now(); const out = obf.obfuscate("ZZc"); const elapsed = performance.now() - start; // Correctness: the search still terminates with a usable, stable redaction. expect(out).toHaveLength(3); expect(obf.obfuscate(out)).toBe(out); expect(obf.obfuscate(obf.obfuscate(out))).toBe(out); // Perf guard: bounded search resolves in ~1ms here; the removed exhaustive // branch took ~70ms+ for this exact case. 30ms is far above the bounded cost // (even generously slower CI hardware) yet well under half the exhaustive // branch's cost, so a reinstated `len <= 3` sweep reliably fails this. expect(elapsed).toBeLessThan(30); }); it("keeps a whole-match default replace regex idempotent when no candidate avoids the regex", () => { // `[\s\S]{8}` matches every possible 8-byte value (unlike `\S{n}` above, it // also matches whitespace and line terminators), so `findNonMatchingReplacement` // exhausts for a value of length >= 3 — the content-hash path the length-1 // sentinel test above cannot reach, since 1-2 char values hardcode to `Z`/`ZZ` // regardless of content. Regression: the whole-match default-replace fallback // kept the content-hash replacement as-is. That replacement is derived from a // hash of the bytes being replaced, so once it is emitted and re-scanned on the // next obfuscate() pass, hashing ITS OWN bytes (not the original secret) // produced a DIFFERENT same-length value — the marker churned forever instead // of reaching a fixed point. The fix falls back to the key+length-only marker, // which is stable because it depends on nothing the marker's own regeneration changes. const obf = new SecretObfuscator([{ type: "regex", mode: "replace", content: "[\\s\\S]{8}" }], "Q".repeat(43)); const out = obf.obfuscate("SECRET12"); expect(out).toHaveLength(8); expect(obf.obfuscate(out)).toBe(out); expect(obf.obfuscate(obf.obfuscate(out))).toBe(out); }); it("keeps a whole-match default replace regex stable across restart when no candidate avoids the regex", () => { // Cross-restart counterpart to the previous test. The key+length-only fallback // marker depends only on the persisted key and the value's length, so a fresh // obfuscator with the same key reproduces the IDENTICAL marker and persisted // text never drifts. Regression: the old content-hash fallback depended on the // bytes of the already-redacted marker, so a fresh obfuscator re-hashed those // marker bytes into a different value, drifting persisted obfuscated text — and // the provider prompt-cache prefix it anchors — across every restart. const key = "R".repeat(43); const entries = [{ type: "regex" as const, mode: "replace" as const, content: "[\\s\\S]{8}" }]; const first = new SecretObfuscator(entries, key); const persisted = first.obfuscate("SECRET12"); const restarted = new SecretObfuscator(entries, key); expect(restarted.obfuscate(persisted)).toBe(persisted); }); it("redacts a context-sensitive replace regex to a value stable in place", () => { // A replace-mode regex with lookbehind matches a value only in context, so a // candidate tested in isolation can be accepted yet re-match once substituted // back. Regression: `(?<=api=)[AZ]` redacting `api=Z` to `api=A` (A does not // match a bare `A`) was then re-redacted to `api=Z` on the next pass, shipping // the raw matched value on every other turn. The replacement must be a fixed // point evaluated in its surrounding context. Both inputs exercise a distinct // path: `Z` is the sentinel collision, `A`'s deterministic replacement (`Z`) // differs from the value but still re-matches in context. for (const input of ["api=Z", "api=A"]) { const obf = new SecretObfuscator( [{ type: "regex", mode: "replace", content: "(?<=api=)[AZ]" }], "Q".repeat(43), ); const out = obf.obfuscate(input); expect(out).toHaveLength(input.length); expect(out.startsWith("api=")).toBe(true); // The matched character must not survive in a position the regex re-matches. expect(/(?<=api=)[AZ]/.test(out)).toBe(false); // Re-obfuscation is a fixed point, so it never oscillates back to the raw value. expect(obf.obfuscate(out)).toBe(out); expect(obf.obfuscate(obf.obfuscate(out))).toBe(out); } }); it("redacts a wide-lookbehind replace regex to a value stable in place", () => { // A lookbehind wider than the re-match back-scan window (512) must still // evaluate against the full surrounding text. Regression: truncating the // fixed-point probe to a fixed window dropped the 600-char lookbehind, so the // check falsely accepted a candidate the regex DOES re-match once the whole // prefix is present — the redaction then oscillated back to the raw matched // value on alternating obfuscate() passes, leaking it to the provider. const prefix = "A".repeat(600); const re = new RegExp(`(?<=${prefix})[AZ]`); for (const input of [`${prefix}Z`, `${prefix}A`]) { const obf = new SecretObfuscator( [{ type: "regex", mode: "replace", content: `(?<=${prefix})[AZ]` }], "Q".repeat(43), ); const out = obf.obfuscate(input); expect(out).toHaveLength(input.length); expect(out.startsWith(prefix)).toBe(true); // The matched character must not survive where the full-context regex re-matches. expect(re.test(out)).toBe(false); // Re-obfuscation is a fixed point, so it never oscillates back to the raw value. expect(obf.obfuscate(out)).toBe(out); expect(obf.obfuscate(obf.obfuscate(out))).toBe(out); } }); it("does not require a placeholder key for entries that never produce a reversible placeholder", () => { // Short plain obfuscate entries are toned down, so they must not force key // creation; regex/long-plain obfuscate entries can placehold and do need it. // A plain replace entry's replacement is pure content-hash, so it never needs // the key either — see the tests below for regex replace entries, which can // still need the key for a different reason (the fixed-point fallback marker). expect(secretEntryNeedsPlaceholderKey({ type: "plain", content: "abc" })).toBe(false); expect(secretEntryNeedsPlaceholderKey({ type: "plain", content: "abcdefgh" })).toBe(true); expect(secretEntryNeedsPlaceholderKey({ type: "regex", content: "[A-Z]{2}" })).toBe(true); expect(secretEntryNeedsPlaceholderKey({ type: "plain", content: "abc", mode: "replace" })).toBe(false); }); it("requires a placeholder key for a default replace regex but not one with a custom replacement", () => { // A replace-mode regex with NO custom `replacement` can reach // `#generateRegexReplacement`'s no-stable-candidate fallback (e.g. for a // pathological match-everything regex like `[\s\S]{8}`), which derives its // marker from the persisted key so it stays a fixed point across restarts — // so this shape DOES need the key. Regression: this previously returned // `false` for every replace-mode entry, so a fresh install with only this // entry never persisted a key and `SecretObfuscator` fell back to // `defaultPlaceholderKey()` (process-random, regenerated every restart), // churning the fallback marker across restarts anyway — the exact symptom // the fixed-point fix was meant to eliminate. expect(secretEntryNeedsPlaceholderKey({ type: "regex", content: "[\\s\\S]{8}", mode: "replace" })).toBe(true); // A regex WITH a custom `replacement` never reaches that fallback — it always // emits the literal configured string — so it still does not need the key. expect( secretEntryNeedsPlaceholderKey({ type: "regex", content: "[\\s\\S]{8}", mode: "replace", replacement: "REDACTED", }), ).toBe(false); }); it("requires a persisted placeholder key for a config with only a whole-match default replace regex", () => { // End-to-end counterpart to the previous test, exercised through // `secretEntriesNeedPlaceholderKey` — what `sdk.ts` actually calls to decide // whether to create/read the persisted key file. Entries mirror "keeps a // whole-match default replace regex stable across restart" above. Regression: // this previously returned `false` for this exact config, so `sdk.ts` never // persisted a key and `SecretObfuscator` fell back to a process-random key, // churning the fixed-point fallback marker across restarts anyway. const entries = [{ type: "regex" as const, mode: "replace" as const, content: "[\\s\\S]{8}" }]; expect(secretEntriesNeedPlaceholderKey(entries)).toBe(true); }); it("ignores obfuscate entries shadowed by a same-content replace entry when deciding key need", () => { const secret = "ghp_exampletoken1234567890"; // A same-content plain replace entry runs before the plain obfuscate entry in // obfuscate(), so the value is one-way replaced and the obfuscate entry never // emits a reversible placeholder. The set must therefore not require the key. expect( secretEntriesNeedPlaceholderKey([ { type: "plain", content: secret, mode: "obfuscate" }, { type: "plain", content: secret, mode: "replace" }, ]), ).toBe(false); // Verify the shadowing actually holds: no reversible placeholder is emitted. const obf = new SecretObfuscator( [ { type: "plain", content: secret, mode: "obfuscate" }, { type: "plain", content: secret, mode: "replace" }, ], "test-placeholder-key", ); const out = obf.obfuscate(`value=${secret}`); expect(out).not.toContain(secret); expect(out).not.toMatch(/\$\$[A-Z0-9]/); // An unshadowed obfuscate entry still requires the key. expect(secretEntriesNeedPlaceholderKey([{ type: "plain", content: secret, mode: "obfuscate" }])).toBe(true); // A replace entry with DIFFERENT content does not shadow the obfuscate entry. expect( secretEntriesNeedPlaceholderKey([ { type: "plain", content: secret, mode: "obfuscate" }, { type: "plain", content: "unrelated-other-secret", mode: "replace" }, ]), ).toBe(true); // A replace entry whose CUSTOM replacement still contains the secret does NOT // shadow it: obfuscate()'s later pass re-scans the inserted replacement text and // emits a reversible placeholder, which needs the persisted key. const reintroEntries: SecretEntry[] = [ { type: "plain", content: secret, mode: "obfuscate" }, { type: "plain", content: secret, mode: "replace", replacement: `PREFIX_${secret}_SUFFIX` }, ]; expect(secretEntriesNeedPlaceholderKey(reintroEntries)).toBe(true); const reintroOut = new SecretObfuscator(reintroEntries, "test-placeholder-key").obfuscate(`value=${secret}`); expect(reintroOut).toMatch(/\$\$[A-Z0-9]/); // Among duplicate same-content replace entries the LAST one wins (the // obfuscator stores replace mappings in a content-keyed Map), so a safe earlier // duplicate must not mask a later reintroducing one: key is still required. const dupReintroLast: SecretEntry[] = [ { type: "plain", content: secret, mode: "obfuscate" }, { type: "plain", content: secret, mode: "replace" }, { type: "plain", content: secret, mode: "replace", replacement: secret }, ]; expect(secretEntriesNeedPlaceholderKey(dupReintroLast)).toBe(true); expect(new SecretObfuscator(dupReintroLast, "test-placeholder-key").obfuscate(`value=${secret}`)).toMatch( /\$\$[A-Z0-9]/, ); // Reverse order: a later safe replacement overrides an earlier reintroducing // one, so the obfuscate entry is shadowed and no key is needed. const dupSafeLast: SecretEntry[] = [ { type: "plain", content: secret, mode: "obfuscate" }, { type: "plain", content: secret, mode: "replace", replacement: `X_${secret}_X` }, { type: "plain", content: secret, mode: "replace" }, ]; expect(secretEntriesNeedPlaceholderKey(dupSafeLast)).toBe(false); expect(new SecretObfuscator(dupSafeLast, "test-placeholder-key").obfuscate(`value=${secret}`)).not.toMatch( /\$\$[A-Z0-9]/, ); // A transitive replace chain that rewrites a safe alias back into the secret // (`SECRET -> ALIAS`, `ALIAS -> SECRET`) reintroduces the value before the // plain-obfuscate pass, so the key is still required. const alias = "ALIAS_TOKEN_XYZ"; const chainReintro: SecretEntry[] = [ { type: "plain", content: secret, mode: "obfuscate" }, { type: "plain", content: secret, mode: "replace", replacement: alias }, { type: "plain", content: alias, mode: "replace", replacement: secret }, ]; expect(secretEntriesNeedPlaceholderKey(chainReintro)).toBe(true); expect(new SecretObfuscator(chainReintro, "test-placeholder-key").obfuscate(`value=${secret}`)).toMatch( /\$\$[A-Z0-9]/, ); // A replacement fragment that joins with adjacent passthrough bytes to form an // obfuscate content (`A -> SEC`, so `ARET12` becomes `SECRET12` during the // replace phase) still emits a reversible placeholder, so the key is required // even though no single replacement contains the whole content. const fragmentJoin: SecretEntry[] = [ { type: "plain", content: "SECRET12", mode: "obfuscate" }, { type: "plain", content: "SECRET12", mode: "replace", replacement: "SAFE" }, { type: "plain", content: "A", mode: "replace", replacement: "SEC" }, ]; expect(secretEntriesNeedPlaceholderKey(fragmentJoin)).toBe(true); expect(new SecretObfuscator(fragmentJoin, "test-placeholder-key").obfuscate("x ARET12 y")).toMatch( /\$\$[A-Z0-9]/, ); // A delete replacement also joins the passthrough bytes on both sides of the // removed token, so it can reconstruct the obfuscate content even though its // replacement output is empty. const deleteJoin: SecretEntry[] = [ { type: "plain", content: "SECRET12", mode: "obfuscate" }, { type: "plain", content: "SECRET12", mode: "replace", replacement: "SAFE" }, { type: "plain", content: "X", mode: "replace", replacement: "" }, ]; expect(secretEntriesNeedPlaceholderKey(deleteJoin)).toBe(true); expect(new SecretObfuscator(deleteJoin, "test-placeholder-key").obfuscate("SECRETX12")).toMatch(/\$\$[A-Z0-9]/); }); it("does not require the key when a later replacement erases a content-forming fragment", () => { // `AA -> SEC` could seed `SECRET12`, but the later (shorter-content) `S -> X` // rewrites every `SEC` into `XEC`, so the replace phase can never leave // `SECRET12` for the plain-obfuscate pass. The key-need check must model that // erasure and stay key-free instead of writing secret-placeholder.key for an // effectively non-placeholding config. const entries: SecretEntry[] = [ { type: "plain", content: "SECRET12", mode: "obfuscate" }, { type: "plain", content: "SECRET12", mode: "replace", replacement: "SAFE" }, { type: "plain", content: "AA", mode: "replace", replacement: "SEC" }, { type: "plain", content: "S", mode: "replace", replacement: "X" }, ]; expect(secretEntriesNeedPlaceholderKey(entries)).toBe(false); // And the obfuscator never reconstructs the secret nor emits a reversible // placeholder, confirming the config is genuinely non-placeholding. const out = new SecretObfuscator(entries, "test-placeholder-key").obfuscate("AARET12 and AART12"); expect(out).not.toContain("SECRET12"); expect(out).not.toMatch(/\$\$[A-Z0-9]/); }); it("does not require the key when a later replacement erases the surrounding context bytes", () => { // `AA -> SEC` could seed `SECRET12` by tiling its `SEC` output against // passthrough `RET12`, but the later `R -> X` rewrites that surrounding `R`, // so a freshly formed `SECRET12` becomes `SECXET12` before the obfuscate pass // — direct `SECRET12` is also shadowed to `SAFE`. The probe must model that // later replacements destroy the required surrounding bytes, not just the // fragment itself, and stay key-free for this effectively non-placeholding // config (otherwise it writes secret-placeholder.key and fails startup in an // unwritable agent config dir). const entries: SecretEntry[] = [ { type: "plain", content: "SECRET12", mode: "obfuscate" }, { type: "plain", content: "SECRET12", mode: "replace", replacement: "SAFE" }, { type: "plain", content: "AA", mode: "replace", replacement: "SEC" }, { type: "plain", content: "R", mode: "replace", replacement: "X" }, ]; expect(secretEntriesNeedPlaceholderKey(entries)).toBe(false); const out = new SecretObfuscator(entries, "test-placeholder-key").obfuscate("AARET12 and SECRET12"); expect(out).not.toContain("SECRET12"); expect(out).not.toMatch(/\$\$[A-Z0-9]/); }); it("redacts a raw sentinel-shaped suffix bridged into a match by a prior placeholder", () => { // A prior-call placeholder followed by RAW text that merely looks like a // deterministic redaction sentinel (`ZZ…`). The default-replace regex matches // only because the deobfuscated placeholder bridges the combined value. The raw // suffix was never emitted by this obfuscator, so it must be redacted rather than // skipped by shape — both for a fixed-width regex (where the suffix does NOT // independently match) and a variable-width one (where it DOES). Either way, // leaving it would leak `ZZZZ` to the provider. for (const content of ["[A-Z0-9]{12}", "[A-Z0-9]{4,12}"]) { const obf = new SecretObfuscator( [ { type: "plain", content: "SECRETUV" }, { type: "regex", mode: "replace", content }, ], "R".repeat(43), ); const token = obf.obfuscate("SECRETUV"); expect(token).toMatch(/^\$\$[A-Z0-9]+:U\$\$$/); const out = obf.obfuscate(`${token}ZZZZ`); expect(out).toContain(token); expect(out).not.toContain("ZZZZ"); } }); it("keeps default replace regex output idempotent around prior placeholders", () => { const obf = new SecretObfuscator( [ { type: "plain", content: "SECRETUV" }, { type: "regex", mode: "replace", content: "[A-Za-z0-9]{10}" }, ], "M".repeat(43), ); const out = obf.obfuscate("SECRETUVX1"); expect(obf.obfuscate(out)).toBe(out); }); it("keeps broad default replace regex output idempotent around prior placeholders", () => { const obf = new SecretObfuscator( [ { type: "plain", content: "SECRETUV" }, { type: "regex", mode: "replace", content: "[A-Za-z0-9]+" }, ], "O".repeat(43), ); const out = obf.obfuscate("SECRETUVX1"); expect(obf.obfuscate(out)).toBe(out); }); it("keeps default replace regex output idempotent after restart", () => { const key = "P".repeat(43); const first = new SecretObfuscator( [ { type: "plain", content: "SECRETUV" }, { type: "regex", mode: "replace", content: "[A-Za-z0-9]+" }, ], key, ); const persisted = first.obfuscate("SECRETUVX1"); const restarted = new SecretObfuscator( [ { type: "plain", content: "SECRETUV" }, { type: "regex", mode: "replace", content: "[A-Za-z0-9]+" }, ], key, ); expect(restarted.obfuscate(persisted)).toBe(persisted); }); it("keeps a multi-char default replace remainder stable across restart", () => { // Regression: a per-chunk remainder longer than the 2-char `ZZ` sentinel was // redacted to a content-derived `ZZ`+hash chunk that was only a fixed point // within the generating session (tracked in #generatedReplaceChunks). A fresh // obfuscator with the same key re-redacted it to a different value, so persisted // obfuscated text — and the provider prompt-cache prefixes it anchors — drifted // across restart (`ZZPL#…#` -> `ZZ7f#…#`). The remainder marker now depends only // on the key and length, so a fresh instance reproduces it byte-identically. const key = "Z".repeat(43); const entries = [ { type: "plain" as const, content: "ABCDEFGH" }, { type: "regex" as const, mode: "replace" as const, content: "[A-Z0-9]{12}" }, ]; const first = new SecretObfuscator(entries, key); const token = first.obfuscate("ABCDEFGH"); const persisted = first.obfuscate("BBBBABCDEFGH"); // The 4-char remainder is redacted (raw bytes gone) but the placeholder survives. expect(persisted).not.toContain("BBBB"); expect(persisted).toContain(token); expect(first.obfuscate(persisted)).toBe(persisted); // A fresh obfuscator (same key) reprocessing the persisted text must not drift. const restarted = new SecretObfuscator(entries, key); expect(restarted.obfuscate(persisted)).toBe(persisted); }); it("ignores regex matches that fall entirely inside known placeholders", () => { const obfuscator = new SecretObfuscator([ { type: "plain", content: "abcdefgh" }, { type: "regex", mode: "replace", content: "P+", replacement: "REDACTED" }, ]); const obfuscated = obfuscator.obfuscate("abcdefgh"); expect(obfuscated).not.toBe("REDACTED"); expect(obfuscated).toMatch(/^\$\$[A-Z0-9]+:L\$\$$/); expect(obfuscator.deobfuscate(obfuscated)).toBe("abcdefgh"); }); it("ignores obfuscate regex matches that fall entirely inside known placeholders", () => { const obfuscator = new SecretObfuscator([ { type: "plain", content: "abcdefgh" }, { type: "regex", content: "P{8}", friendlyName: "inner" }, ]); const obfuscated = obfuscator.obfuscate("abcdefgh"); expect(obfuscated).toMatch(/^\$\$[A-Z0-9]+:L\$\$$/); expect(obfuscated).not.toMatch(/^\$\$INNER_/); expect(obfuscator.deobfuscate(obfuscated)).toBe("abcdefgh"); }); it("ignores obfuscate regex matches that partially overlap known placeholders", () => { const obfuscator = new SecretObfuscator([ { type: "plain", content: "secretuv" }, { type: "regex", content: "P{3}X", friendlyName: "partial" }, ]); const obfuscated = obfuscator.obfuscate("secretuvX"); expect(obfuscated).toMatch(/^\$\$[A-Z0-9]+:L\$\$X$/); expect(obfuscated).not.toMatch(/^\$\$PARTIAL_/); expect(obfuscator.deobfuscate(obfuscated)).toBe("secretuvX"); }); it("does not recursively rewrite plain secrets that look like placeholders", () => { const sharedKey = "D".repeat(43); const firstOnly = new SecretObfuscator( [{ type: "plain", content: "legacy-secret", friendlyName: "old" }], sharedKey, ); const firstPlaceholder = firstOnly.obfuscate("legacy-secret"); const secondOnly = new SecretObfuscator( [{ type: "plain", content: firstPlaceholder, friendlyName: "other" }], sharedKey, ); const secondPlaceholder = secondOnly.obfuscate(firstPlaceholder); const obfuscator = new SecretObfuscator( [ { type: "plain", content: "legacy-secret", friendlyName: "old" }, { type: "plain", content: firstPlaceholder, friendlyName: "other" }, ], sharedKey, ); expect(secondPlaceholder).toMatch(/^\$\$OTHER_[A-Z0-9]+(?::[ULCM])?\$\$$/); expect(obfuscator.deobfuscate(secondPlaceholder)).toBe(firstPlaceholder); }); it("keeps no-name placeholders unprefixed but content-derived", () => { const first = new SecretObfuscator([ { type: "plain", content: "alpha-secret" }, { type: "plain", content: "beta-secret" }, ]); const second = new SecretObfuscator([ { type: "plain", content: "beta-secret" }, { type: "plain", content: "alpha-secret" }, ]); const firstToken = first.obfuscate("alpha-secret").match(/\$\$[A-Z0-9]+:L\$\$/)?.[0]; const secondToken = second.obfuscate("alpha-secret").match(/\$\$[A-Z0-9]+:L\$\$/)?.[0]; expect(firstToken).toBeDefined(); expect(firstToken).toBe(secondToken); expect(firstToken).not.toMatch(/_[A-Z0-9]+/); expect(first.deobfuscate(firstToken ?? "")).toBe("alpha-secret"); }); it("leaves hash-delimited tokens inert while restoring current placeholders", () => { const obfuscator = new SecretObfuscator([{ type: "plain", content: "legacy-secret" }]); const current = obfuscator.obfuscate("legacy-secret"); const message: AgentMessage = { role: "assistant", content: [{ type: "text", text: `legacy #XRRS# current ${current}` }], api: "anthropic-messages", provider: "anthropic", model: "test-model", usage: { input: 1, output: 1, cacheRead: 0, cacheWrite: 0, totalTokens: 2, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 }, }, stopReason: "stop", timestamp: 1, }; const restored = deobfuscateAgentMessages(obfuscator, [message])[0]; if (restored?.role !== "assistant") throw new Error("expected restored assistant message"); const text = restored.content[0]; expect(text?.type === "text" && text.text).toBe("legacy #XRRS# current legacy-secret"); }); it("deobfuscates placeholders after friendlyName changes", () => { const renamed = new SecretObfuscator([{ type: "plain", content: "renamed-secret", friendlyName: "new name" }]); const current = renamed.obfuscate("renamed-secret"); const oldName = current.replace("$$NEWNAME_", () => "$$OLDNAME_"); const removedName = new SecretObfuscator([{ type: "plain", content: "renamed-secret" }]); expect(current).toMatch(/^\$\$NEWNAME_[A-Z0-9]+:L\$\$$/); expect(renamed.deobfuscate(oldName)).toBe("renamed-secret"); expect(removedName.deobfuscate(oldName)).toBe("renamed-secret"); }); it("keeps friendly-name-independent aliases unique for same-base same-hint secrets", () => { const obfuscator = new SecretObfuscator([ { type: "plain", content: "SeCretuv", friendlyName: "alpha" }, { type: "plain", content: "SecRetuv", friendlyName: "bravo" }, ]); const obfuscated = obfuscator.obfuscate("SeCretuv SecRetuv"); const [tokenA, tokenB] = obfuscated.split(" "); if (!tokenA || !tokenB) throw new Error("expected two friendly placeholders"); expect(tokenA).toMatch(/^\$\$ALPHA_[A-Z0-9]+:M\$\$$/); expect(tokenB).toMatch(/^\$\$BRAVO_[A-Z0-9]+:M\$\$$/); expect(obfuscator.deobfuscate(obfuscated)).toBe("SeCretuv SecRetuv"); const aliasA = tokenA.replace(/^\$\$[A-Z0-9]+_/, () => "$$"); const aliasB = tokenB.replace(/^\$\$[A-Z0-9]+_/, () => "$$"); expect(aliasA).not.toBe(aliasB); expect(obfuscator.deobfuscate(aliasA)).toBe("SeCretuv"); expect(obfuscator.deobfuscate(aliasB)).toBe("SecRetuv"); }); it("resolves a persisted friendly placeholder to the right same-base secret after a rename", () => { const original = new SecretObfuscator([ { type: "plain", content: "SeCretuv", friendlyName: "alpha" }, { type: "plain", content: "SecRetuv", friendlyName: "bravo" }, ]); const persistedBravo = original.obfuscate("SeCretuv SecRetuv").split(" ")[1]; if (!persistedBravo) throw new Error("expected a bravo placeholder"); // bravo renamed to charlie while both same-base secrets still exist. const renamed = new SecretObfuscator([ { type: "plain", content: "SeCretuv", friendlyName: "alpha" }, { type: "plain", content: "SecRetuv", friendlyName: "charlie" }, ]); expect(renamed.deobfuscate(persistedBravo)).toBe("SecRetuv"); }); it("rejects a forged placeholder wrapper that smuggles a real secret prefix past obfuscate()", () => { // Regression: #isGeneratedPlaceholder decides "is this span already a // safely-generated placeholder" for a friendly-prefixed candidate // (`#PREFIX_HASH:HINT#`) by falling back to the friendly-name-independent // bare alias `#HASH:HINT#` — the SAME loose lookup deobfuscate() // intentionally uses so a mangled/renamed friendly-name prefix still // round-trips. Accepting ANY prefix whose bare alias resolves let // untrusted text forge `#_:#`: // the other secret's bare alias resolves fine (it really is configured), // so the WHOLE forged token — including the exposed secret literal // standing in for the friendly name — was treated as already-redacted and // passed through obfuscate() untouched. The fix refuses the bare-alias // fallback whenever the dropped prefix CONTAINS a configured secret's // literal value (exactly, or as a substring), so a forged prefix built // from a real secret is never treated as safe, while a genuine // friendly-name label (or a stale one after a rename) still is. const secretA = "LEAKEDSECRETALPHA"; // uppercase-alnum: fits the placeholder-prefix grammar const secretB = "bravo-secret-9f3d8c2b"; const obfuscator = new SecretObfuscator([ { type: "plain", content: secretA, friendlyName: "ALPHA" }, { type: "plain", content: secretB, friendlyName: "BRAVO" }, ]); // Learn secretB's real generated placeholder, then derive the friendly- // name-independent bare-alias suffix (`_:#`) that // lookupFriendlyPlaceholderAlias also resolves on purpose for deobfuscate(). const bravoPlaceholder = obfuscator.obfuscate(secretB); expect(bravoPlaceholder).toMatch(/^\$\$BRAVO_[A-Z0-9]{4,}(?::[ULCM])?\$\$$/); const aliasSuffix = bravoPlaceholder.replace(/^\$\$BRAVO/, ""); // Forge a token shaped exactly like a friendly placeholder for secretB, but // with secretA's raw literal value standing in for the friendly name. const forgedExact = `$$${secretA}${aliasSuffix}`; expect(forgedExact).toMatch(/^\$\$[A-Z0-9]+_[A-Z0-9]{4,}(?::[ULCM])?\$\$$/); expect(obfuscator.obfuscate(forgedExact)).not.toContain(secretA); // A prefix that merely CONTAINS secretA (not just equals it) must also be // rejected — the fix scans for containment, not exact equality. const forgedSubstring = `#X${secretA}Y${aliasSuffix}`; expect(obfuscator.obfuscate(forgedSubstring)).not.toContain(secretA); // Mixed real + forged in one call: the real secretB placeholder must still // round-trip normally via deobfuscate(), and the forged wrapper must still // not leak secretA. const mixed = `${bravoPlaceholder} then ${forgedExact}`; const mixedObfuscated = obfuscator.obfuscate(mixed); expect(mixedObfuscated).not.toContain(secretA); expect(mixedObfuscated).toContain(bravoPlaceholder); expect(obfuscator.deobfuscate(bravoPlaceholder)).toBe(secretB); }); it("rejects a forged placeholder wrapper that smuggles a regex-discovered secret prefix past obfuscate()", () => { // Regression: the forged-prefix check above only ever scanned // `#configuredSecretValues` (secrets known up front). A regex-discovered // secret is never in that set — regex secrets are found dynamically as // obfuscate() encounters them — so a REGEX secret's literal value could // still be wrapped as the forged "friendly name" prefix around another // secret's real bare-alias suffix and sail through #isGeneratedPlaceholder // untouched. The fix also scans every value this instance has ever minted // an obfuscate-mode placeholder for (`#obfuscateMappings`, which regex // discoveries populate too), not just the always-known configured set. const regexSecret = "LEAKTOKEN42"; const secretB = "bravo-secret-value"; const obfuscator = new SecretObfuscator([ { type: "regex", content: "LEAKTOKEN[0-9]+", friendlyName: "leak" }, { type: "plain", content: secretB, friendlyName: "BRAVO" }, ]); // Force the regex secret to be discovered/minted first, matching the real // attack model of "attacker has observed this secret's placeholder in an // earlier turn" — only after this does #obfuscateMappings know its value. const regexPlaceholder = obfuscator.obfuscate(regexSecret); expect(regexPlaceholder).not.toContain(regexSecret); // Learn secretB's real placeholder, then derive the friendly-name- // independent bare-alias suffix (`_:#`). const bravoPlaceholder = obfuscator.obfuscate(secretB); expect(bravoPlaceholder).toMatch(/^\$\$BRAVO_[A-Z0-9]{4,}(?::[ULCM])?\$\$$/); const aliasSuffix = bravoPlaceholder.replace(/^\$\$BRAVO/, ""); // Forge a token wrapping the REGEX secret's raw literal around secretB's // real bare-alias suffix. const forgedExact = `$$${regexSecret}${aliasSuffix}`; expect(forgedExact).toMatch(/^\$\$[A-Z0-9]+_[A-Z0-9]{4,}(?::[ULCM])?\$\$$/); expect(obfuscator.obfuscate(forgedExact)).not.toContain(regexSecret); // Mixed real + forged in one call: the real secretB placeholder must still // round-trip normally via deobfuscate(), and the forged wrapper must still // not leak the regex secret. const mixed = `${bravoPlaceholder} then ${forgedExact}`; const mixedObfuscated = obfuscator.obfuscate(mixed); expect(mixedObfuscated).not.toContain(regexSecret); expect(mixedObfuscated).toContain(bravoPlaceholder); expect(obfuscator.deobfuscate(bravoPlaceholder)).toBe(secretB); }); it("rejects a forged alias wrapper whose prefix only matches a case-variant regex-discovered secret", () => { // Regression: the forged-prefix checks above only ever scanned EXACT // previously-DISCOVERED secret strings (`#configuredSecretValues` and // `#obfuscateMappings`), recorded in whatever casing they first turned // up in. A case-insensitive (or other flag-variant) regex — e.g. // `{ type: "regex", content: "tok[a-z0-9]+", flags: "i" }` — only ever // records the ONE casing it actually discovered (lowercase // `tokabc123`) in `#obfuscateMappings`. A forged token wrapping a // DIFFERENTLY-cased occurrence of that same secret-shaped text // (`TOKABC123`) around a real bare-alias suffix matched neither exact- // string check, so it sailed through #isGeneratedPlaceholder as an // already-generated placeholder and the uppercase secret text leaked // verbatim. The fix also tests the dropped prefix directly against the // configured regex's own pattern, which matches regardless of the // casing the secret was first discovered under. const obf = new SecretObfuscator([{ type: "regex", content: "tok[a-z0-9]+", flags: "i" }], "Q".repeat(43)); // Discover the secret in lowercase, minting a real bare placeholder and // registering `tokabc123` (not `TOKABC123`) in `#obfuscateMappings`. const real = obf.obfuscate("use tokabc123 now"); expect(real).not.toContain("tokabc123"); const suffix = /\$\$([A-Z0-9]{4,}(?::[ULCM])?)\$\$/.exec(real)?.[1]; expect(suffix).toBeDefined(); // Forge a token wrapping an UPPERCASE variant of the secret around the // real bare-alias suffix — differently cased from what was actually // discovered, so it cannot match either exact-string check, only the // regex pattern itself. const forged = `see $$TOKABC123_${suffix}$$ here`; const out = obf.obfuscate(forged); expect(out).not.toContain("TOKABC123"); }); it("refuses deobfuscate()'s bare-alias fallback for a forged secret-shaped prefix, but still honors a genuine rename", () => { // Regression: `#lookupLiveAlias`'s bare-alias fallback (`#PREFIX_HASH#` → // strip prefix → look up bare `#HASH#`) used to accept ANY prefix // unconditionally as long as the bare hash suffix belonged to a real // placeholder. On the live provider-output / tool-call-argument restore // path, an attacker who observed any real placeholder's bare hash suffix // elsewhere in a transcript could wrap it in a forged prefix that is a // sanitized/normalized rendering of a configured secret's own value (or a // different secret's), and deobfuscate() would restore that OTHER // secret's raw value in its place — worse than the obfuscate-direction // leak defended against above, since this is the path that reconstitutes // real secrets for outbound tool calls. The fix reuses // `#prefixIsSecretShaped` (shared with `#isGeneratedPlaceholder`) to // refuse the fallback whenever the dropped prefix is itself // secret-shaped, while a genuine friendly-name rename — a prefix that // matches no configured secret value or pattern — still round-trips. const secret = "github_pat_abc123"; const obfuscator = new SecretObfuscator([{ type: "plain", content: secret }]); const real = obfuscator.obfuscate(secret); const suffix = /\$\$([A-Z0-9]{4,}(?::[ULCM])?)\$\$/.exec(real)?.[1]; expect(suffix).toBeDefined(); // Forge a prefix that is the sanitized rendering of the SAME configured // secret's own value, wrapped around the real bare-alias suffix. const forged = `run tool with $$GITHUBPATABC123_${suffix}$$ now`; const restored = obfuscator.deobfuscate(forged); expect(restored).not.toContain(secret); expect(restored).toContain("$$GITHUBPATABC123_"); // A genuine rename is unaffected: the OLD friendly-name prefix is not // itself secret-shaped, so the bare-alias fallback still resolves it to // the same secret under the RENAMED (equally non-secret-shaped) prefix. const renameObfuscator = new SecretObfuscator([ { type: "plain", content: "some-other-secret-value", friendlyName: "OldName" }, ]); const currentToken = renameObfuscator.obfuscate("some-other-secret-value"); expect(currentToken).toMatch(/^\$\$OLDNAME_[A-Z0-9]+:L\$\$$/); const renameSuffix = currentToken.replace(/^\$\$OLDNAME/, ""); expect(renameObfuscator.deobfuscate(`$$NEWNAME${renameSuffix}`)).toBe("some-other-secret-value"); }); it("drops a friendly name that contains another configured secret's literal value", () => { // Regression: a friendlyName is baked verbatim into every placeholder // minted for its secret (`#NAME_hash:hint#`) via an EXACT // `#deobfuscateMap` entry — not through the loose alias-fallback lookup // the forged-wrapper tests above defend against. If that friendly name // happens to CONTAIN another LIVE secret's literal value, the baked-in // prefix leaks that other secret on every use of THIS secret — and, // because the whole token is now a recognized "already generated" // placeholder, the leaked literal is protected from ever being redacted // by that other secret's own plain-secret pass. The fix refuses the // friendly-name prefix for this specific mint (falling back to a bare, // unprefixed placeholder) whenever the sanitized name contains a // configured plain secret's literal value. const obfuscator = new SecretObfuscator([ { type: "plain", content: "ABCDEFGH", friendlyName: "LEAKTOKEN" }, { type: "plain", content: "LEAKTOKEN" }, { type: "plain", content: "OTHERSECRETXY", friendlyName: "SAFE" }, ]); const obfuscated = obfuscator.obfuscate("ABCDEFGH"); expect(obfuscated).not.toContain("LEAKTOKEN"); // Friendly name dropped for this mint: bare, unprefixed placeholder shape. expect(obfuscated).toMatch(/^\$\$[A-Z0-9]{4,}(?::[ULCM])?\$\$$/); // A friendly name that does NOT collide with any live secret is unaffected. const safeObfuscated = obfuscator.obfuscate("OTHERSECRETXY"); expect(safeObfuscated).toMatch(/^\$\$SAFE_[A-Z0-9]{4,}(?::[ULCM])?\$\$$/); }); it("drops a friendly name matched by a later-declared regex secret, regardless of entries[] order", () => { // Regression: the collision guard above must also cover configured REGEX // patterns, not just plain-secret literals — and must do so regardless of // where the regex entry sits in `entries[]`. The constructor compiles // every regex entry into `#regexEntries` in a dedicated pass BEFORE any // placeholder is minted, so a plain entry's friendly name is checked // against a fully-populated regex set even when its colliding regex entry // is declared LATER in the array. A naive fix that only pre-collected // plain-secret literals (leaving regex compilation in the same forward // pass that mints placeholders) would miss exactly this ordering: at mint // time for the entry below, `#regexEntries` would still be empty. const obfuscator = new SecretObfuscator([ { type: "plain", content: "ABCDEFGH", friendlyName: "LEAKTOKEN" }, { type: "regex", content: "LEAKTOKEN" }, ]); const obfuscated = obfuscator.obfuscate("ABCDEFGH"); expect(obfuscated).not.toContain("LEAKTOKEN"); expect(obfuscated).toMatch(/^\$\$[A-Z0-9]{4,}(?::[ULCM])?\$\$$/); }); it("keeps a mixed-case placeholder stable when a same-normalized secret is added earlier", () => { // Session 1: only SecRet is configured; persist its mixed-case token. const before = new SecretObfuscator([{ type: "plain", content: "SecRetuv" }]); const persisted = before.obfuscate("SecRetuv"); expect(persisted).toMatch(/^\$\$[A-Z0-9]+:M\$\$$/); // Session 2: SeCret (same normalized value, also :M) is added EARLIER. const after = new SecretObfuscator([ { type: "plain", content: "SeCretuv" }, { type: "plain", content: "SecRetuv" }, ]); // SecRet's token must be value-stable, not bumped to a fallback, so the // persisted placeholder still round-trips to SecRet rather than SeCret. expect(after.obfuscate("SecRetuv")).toBe(persisted); expect(after.deobfuscate(persisted)).toBe("SecRetuv"); expect(after.obfuscate("SeCretuv")).not.toBe(persisted); }); it("derives each placeholder purely from its own secret, independent of load order", () => { // A secret persisted alone must keep the same token when unrelated secrets // are later added before it, so old session text never aliases to another // secret because of config/env ordering. const alone = new SecretObfuscator([{ type: "plain", content: "secret397" }]); const persisted = alone.obfuscate("secret397"); const before = new SecretObfuscator([ { type: "plain", content: "secret658" }, { type: "plain", content: "secret397" }, ]); const after = new SecretObfuscator([ { type: "plain", content: "secret397" }, { type: "plain", content: "secret658" }, ]); expect(before.obfuscate("secret397")).toBe(persisted); expect(after.obfuscate("secret397")).toBe(persisted); expect(before.deobfuscate(persisted)).toBe("secret397"); expect(before.obfuscate("secret658")).not.toBe(persisted); }); it("keeps Unicode case variants on distinct bases despite a shared ASCII hint", () => { // `Äbc` and `äbc` differ only by Unicode case; the ASCII case hint (`:L`) // cannot reconstruct Unicode casing, so they must NOT share a base key. A // `secret.toLowerCase()` normalization folds `Ä`→`ä` and collapses them, // letting a persisted token alias to whichever secret loads first. const alone = new SecretObfuscator([{ type: "plain", content: "Äbcdefgh" }]); const persisted = alone.obfuscate("Äbcdefgh"); // A later session loads `äbc` EARLIER than `Äbc`. const reordered = new SecretObfuscator([ { type: "plain", content: "äbcdefgh" }, { type: "plain", content: "Äbcdefgh" }, ]); expect(reordered.obfuscate("Äbcdefgh")).toBe(persisted); expect(reordered.obfuscate("äbcdefgh")).not.toBe(persisted); expect(reordered.deobfuscate(persisted)).toBe("Äbcdefgh"); expect(reordered.deobfuscate(reordered.obfuscate("äbcdefgh"))).toBe("äbcdefgh"); }); it("derives placeholders from a keyed digest, not a public content hash", () => { // A provider that sees the placeholder and knows the algorithm must not be // able to dictionary low-entropy secrets: the base is keyed by a private // per-install secret, so the same secret yields different tokens per key. const secret = "hunter2-password"; const keyA = new SecretObfuscator([{ type: "plain", content: secret }], "install-key-a"); const keyB = new SecretObfuscator([{ type: "plain", content: secret }], "install-key-b"); const tokenA = keyA.obfuscate(secret); const tokenB = keyB.obfuscate(secret); expect(tokenA).not.toBe(secret); expect(tokenA).not.toBe(tokenB); // Same key + same secret is stable across instances and round-trips. const keyAgain = new SecretObfuscator([{ type: "plain", content: secret }], "install-key-a"); expect(keyAgain.obfuscate(secret)).toBe(tokenA); expect(keyA.deobfuscate(tokenA)).toBe(secret); }); it("redacts its own keyed-hash key from obfuscated output", () => { // The key can be read from a user-readable config file by a prompt-injected // tool; if it reached the provider verbatim, the keyed placeholder bases // could be dictionaried, so the obfuscator must never emit its own key. const key = "install-key-that-must-never-leak-1234567890"; const obfuscator = new SecretObfuscator([{ type: "plain", content: "real-secret" }], key); const obfuscated = obfuscator.obfuscate(`cat secret-placeholder.key => ${key} (and real-secret)`); expect(obfuscated).not.toContain(key); expect(obfuscated).not.toContain("real-secret"); }); it("withholds pending placeholders while streaming provider text", () => { expect(stripPendingSecretPlaceholderSuffix("before $")).toBe("before "); expect(stripPendingSecretPlaceholderSuffix("before $$")).toBe("before "); expect(stripPendingSecretPlaceholderSuffix("before $$AB12:")).toBe("before "); expect(stripPendingSecretPlaceholderSuffix("before $$TOKEN")).toBe("before "); expect(stripPendingSecretPlaceholderSuffix("before $$TOKEN_")).toBe("before "); expect(stripPendingSecretPlaceholderSuffix("before $$TOKEN_AB12:")).toBe("before "); expect(stripPendingSecretPlaceholderSuffix("before $$TOKEN_AB12:U")).toBe("before "); // A trailing delimiter character is buffered because it can open an adjacent placeholder. expect(stripPendingSecretPlaceholderSuffix("before $$TOKEN_AB12:U$$")).toBe("before $$TOKEN_AB12:U"); expect(stripPendingSecretPlaceholderSuffix("prefix ID$")).toBe("prefix ID"); expect(stripPendingSecretPlaceholderSuffix("count 42$")).toBe("count 42"); expect(stripPendingSecretPlaceholderSuffix("before $$TOKEN ")).toBe("before $$TOKEN "); // A single `$` followed by non-`$` can never open a `$$…$$` placeholder, so it // must stream through instead of being withheld until the next boundary. expect(stripPendingSecretPlaceholderSuffix("run $HOME")).toBe("run $HOME"); expect(stripPendingSecretPlaceholderSuffix("pay $100")).toBe("pay $100"); expect(stripPendingSecretPlaceholderSuffix("cost $$100")).toBe("cost "); }); it("uses independent bases across casing variants with distinct hints", () => { const obfuscator = new SecretObfuscator([ { type: "plain", content: "secretuv", friendlyName: "token" }, { type: "plain", content: "SECRETUV", friendlyName: "token" }, { type: "plain", content: "Secretuv", friendlyName: "token" }, ]); const obfuscated = obfuscator.obfuscate("secretuv SECRETUV Secretuv"); const tokens = obfuscated.match(/\$\$TOKEN_[A-Z0-9]+:[ULCM]\$\$/g); if (!tokens) throw new Error("Expected case-hinted placeholders"); const bases = tokens.map(token => /^\$\$TOKEN_([A-Z0-9]+):/.exec(token)?.[1]); expect(tokens).toHaveLength(3); // Distinct ASCII-case variants must NOT share a base: a shared case-folded // base would let a provider synthesize a sibling token by swapping the hint. expect(new Set(bases).size).toBe(3); expect(tokens[0]?.endsWith(":L$$")).toBe(true); expect(tokens[1]?.endsWith(":U$$")).toBe(true); expect(tokens[2]?.endsWith(":C$$")).toBe(true); expect(obfuscator.deobfuscate(obfuscated)).toBe("secretuv SECRETUV Secretuv"); }); it("does not restore a case-variant sibling synthesized by swapping the hint", () => { // P1: two obfuscate-mode secrets differing only by ASCII case. Only the // lowercase one is ever provider-visible; a prompt-injected model must not // recover the uppercase secret (never emitted) by taking the visible // token's base and swapping the case hint in a tool-call argument. const key = "case-variant-install-key-0000000000000000000"; const obfuscator = new SecretObfuscator( [ { type: "plain", content: "abc12345" }, { type: "plain", content: "ABC12345" }, ], key, ); // Provider sees only the lowercase placeholder. const visible = obfuscator.obfuscate("abc12345"); expect(visible).toMatch(/^\$\$[A-Z0-9]+:L\$\$$/); const base = /^\$\$([A-Z0-9]+):L\$\$$/.exec(visible)?.[1]; if (!base) throw new Error("expected a lowercase placeholder base"); // The uppercase secret's real token uses an independent base. const upperReal = obfuscator.obfuscate("ABC12345"); expect(upperReal).toMatch(/^\$\$[A-Z0-9]+:U\$\$$/); expect(upperReal).not.toBe(`$$${base}:U$$`); // Live deobfuscation of the synthesized sibling token leaves it literal // instead of restoring the never-provider-visible uppercase secret. const synthesized = `$$${base}:U$$`; expect(obfuscator.deobfuscate(synthesized)).toBe(synthesized); expect(obfuscator.deobfuscateObject({ cmd: synthesized })).toEqual({ cmd: synthesized }); // The legitimate visible token still round-trips. expect(obfuscator.deobfuscate(visible)).toBe("abc12345"); }); it("gives duplicate mixed-case variants distinct placeholders", () => { const obfuscator = new SecretObfuscator([ { type: "plain", content: "SeCretuv", friendlyName: "token" }, { type: "plain", content: "SecRetuv", friendlyName: "token" }, ]); const repeated = new SecretObfuscator([ { type: "plain", content: "SeCretuv", friendlyName: "token" }, { type: "plain", content: "SecRetuv", friendlyName: "token" }, ]); const input = "SeCretuv SecRetuv"; const obfuscated = obfuscator.obfuscate(input); const tokens = obfuscated.match(/\$\$TOKEN_[A-Z0-9]+:M\$\$/g); if (!tokens) throw new Error("Expected mixed-case placeholders"); expect(tokens).toHaveLength(2); expect(new Set(tokens).size).toBe(2); expect(repeated.obfuscate(input)).toBe(obfuscated); expect(obfuscator.deobfuscate(obfuscated)).toBe(input); }); it("allows duplicate friendly names because hash suffixes disambiguate them", () => { const obfuscator = new SecretObfuscator([ { type: "plain", content: "first-token", friendlyName: "api" }, { type: "plain", content: "second-token", friendlyName: "api" }, ]); const obfuscated = obfuscator.obfuscate("first-token second-token"); const tokens = obfuscated.match(/\$\$API_[A-Z0-9]+:L\$\$/g); if (!tokens) throw new Error("Expected friendly-name placeholders"); expect(tokens).toHaveLength(2); expect(new Set(tokens).size).toBe(2); expect(obfuscator.deobfuscate(obfuscated)).toBe("first-token second-token"); }); it("sanitizes and caps friendly names", () => { expect(sanitizeSecretFriendlyName("git hub-token!!!")).toBe("GITHUBTOKEN"); expect(sanitizeSecretFriendlyName("0123456789abcdefghijklmnopqrstuvwxyz")).toBe( "0123456789ABCDEFGHIJKLMNOPQRSTUV", ); expect(sanitizeSecretFriendlyName("***")).toBeUndefined(); }); it("omits invalid friendlyName metadata without dropping the secret", async () => { const root = await fs.mkdtemp(path.join(os.tmpdir(), "omp-secret-friendly-")); try { const project = path.join(root, "project"); const agentDir = path.join(root, "agent"); await fs.mkdir(path.join(project, ".omp"), { recursive: true }); await fs.mkdir(agentDir, { recursive: true }); await fs.writeFile( path.join(project, ".omp", "secrets.yml"), "- type: plain\n content: invalid-friendly-secret\n friendlyName: '***'\n", ); const entries = await loadSecrets(project, agentDir); const obfuscator = new SecretObfuscator(entries); const obfuscated = obfuscator.obfuscate("invalid-friendly-secret"); expect(entries).toHaveLength(1); expect(entries[0]?.friendlyName).toBeUndefined(); expect(obfuscated).toMatch(/\$\$[A-Z0-9]+:L\$\$/); expect(obfuscated).not.toMatch(/_[A-Z0-9]+/); expect(obfuscator.deobfuscate(obfuscated)).toBe("invalid-friendly-secret"); } finally { await fs.rm(root, { recursive: true, force: true }); } }); it("omits non-string friendlyName metadata without dropping the secret", async () => { const root = await fs.mkdtemp(path.join(os.tmpdir(), "omp-secret-friendly-")); try { const project = path.join(root, "project"); const agentDir = path.join(root, "agent"); await fs.mkdir(path.join(project, ".omp"), { recursive: true }); await fs.mkdir(agentDir, { recursive: true }); await fs.writeFile( path.join(project, ".omp", "secrets.yml"), "- type: plain\n content: non-string-friendly-secret\n friendlyName: 123\n", ); const entries = await loadSecrets(project, agentDir); const obfuscator = new SecretObfuscator(entries); const obfuscated = obfuscator.obfuscate("non-string-friendly-secret"); expect(entries).toHaveLength(1); expect(entries[0]?.friendlyName).toBeUndefined(); expect(obfuscated).toMatch(/\$\$[A-Z0-9]+:L\$\$/); expect(obfuscated).not.toMatch(/_[A-Z0-9]+/); expect(obfuscator.deobfuscate(obfuscated)).toBe("non-string-friendly-secret"); } finally { await fs.rm(root, { recursive: true, force: true }); } }); it("rejects a secrets.yml replace regex entry that can never redact a 1-2 char match distinctly from itself", async () => { const root = await fs.mkdtemp(path.join(os.tmpdir(), "omp-secret-friendly-")); try { const project = path.join(root, "project"); const agentDir = path.join(root, "agent"); await fs.mkdir(path.join(project, ".omp"), { recursive: true }); await fs.mkdir(agentDir, { recursive: true }); await fs.writeFile( path.join(project, ".omp", "secrets.yml"), '- type: regex\n mode: replace\n content: "."\n', ); const entries = await loadSecrets(project, agentDir); expect(entries).toEqual([]); } finally { await fs.rm(root, { recursive: true, force: true }); } }); it("preserves the raw friendlyName from secrets.yml so a regex entry's self-collision check still catches it", async () => { // Regression: `loadFriendlyName` used to return the SANITIZED (uppercased, // separator-stripped) friendlyName, which then got stored verbatim as // `entry.friendlyName` in the loaded `SecretEntry`. That defeated the // regex-entry self-collision check above (see "rejects a regex entry // friendlyName that is itself a live match for its own pattern"), which // must test the RAW label against the configured pattern: a // case-sensitive/punctuated pattern like `tok_[a-z0-9]+` can never match // an already-uppercased, separator-stripped rendering of itself. The fix // still validates the friendlyName but returns it unsanitized. const root = await fs.mkdtemp(path.join(os.tmpdir(), "omp-secret-friendly-")); try { const project = path.join(root, "project"); const agentDir = path.join(root, "agent"); await fs.mkdir(path.join(project, ".omp"), { recursive: true }); await fs.mkdir(agentDir, { recursive: true }); await fs.writeFile( path.join(project, ".omp", "secrets.yml"), '- type: regex\n content: "tok_[a-z0-9]+"\n friendlyName: "tok_abc123"\n', ); const entries = await loadSecrets(project, agentDir); expect(entries[0]?.friendlyName).toBe("tok_abc123"); const obfuscator = new SecretObfuscator(entries); const obfuscated = obfuscator.obfuscate("use tok_abc123 now"); expect(obfuscated).not.toMatch(/TOKABC123_/); expect(obfuscated).toMatch(/^use \$\$[A-Z0-9]+:L\$\$ now$/); } finally { await fs.rm(root, { recursive: true, force: true }); } }); }); describe("SecretObfuscator cross-turn cache stability", () => { // The provider prompt cache is content-addressed: convertToLlm / transformProviderContext // re-run obfuscation over the WHOLE message array every turn, so a non-deterministic // placeholder for the same secret would rewrite already-sent prefix bytes and bust the // cache (cacheWrite @ $6.25/M vs cacheRead @ $0.50/M on opus). These tests pin the // determinism that makes obfuscation cache-safe so a future change cannot silently // reintroduce per-turn cache invalidation. it("produces byte-identical output when re-obfuscating the same content across turns", () => { const secret = "SUPER_SECRET_TOKEN_12345"; const obfuscator = new SecretObfuscator([ { type: "plain", content: secret }, { type: "regex", content: "tok_[a-z0-9]+" }, ]); const messages: Message[] = [{ role: "user", content: `use ${secret} and tok_abc123`, timestamp: 1 }]; const turn1 = JSON.stringify(obfuscateMessages(obfuscator, messages)); const turn2 = JSON.stringify(obfuscateMessages(obfuscator, messages)); expect(turn1).not.toContain(secret); expect(turn1).not.toContain("tok_abc123"); // Identical bytes on the second pass → the cached prefix stays valid. expect(turn2).toEqual(turn1); }); it("keeps earlier message placeholders stable when a later message reveals a new regex secret", () => { const obfuscator = new SecretObfuscator([{ type: "regex", content: "tok_[a-z0-9]+" }]); const early: Message[] = [{ role: "user", content: "first uses tok_aaaa", timestamp: 1 }]; // Turn N: only the early message exists; tok_aaa mints a fresh placeholder. const earlyTurnN = JSON.stringify(obfuscateMessages(obfuscator, early)); expect(earlyTurnN).not.toContain("tok_aaaa"); // A later turn reveals a brand-new secret. Lazy regex discovery assigns it a fresh // index — this MUST NOT shift the placeholder already minted for tok_aaa. const later: Message[] = [{ role: "user", content: "later uses tok_bbbb", timestamp: 2 }]; const laterOut = JSON.stringify(obfuscateMessages(obfuscator, later)); expect(laterOut).not.toContain("tok_bbbb"); // Re-obfuscate the early message after the new discovery: identical bytes → the // already-cached prefix for the early message stays valid. const earlyTurnNPlus1 = JSON.stringify(obfuscateMessages(obfuscator, early)); expect(earlyTurnNPlus1).toEqual(earlyTurnN); }); it("shares regex-protected values across the whole outbound batch so an earlier message's friendly prefix cannot leak a later message's secret", () => { // Regression: obfuscateMessages precomputes `sharedRegexSecretValues` across // ALL targeted message texts before obfuscating any single one. Without that // precomputation, processing messages one by one would let an EARLIER message // mint a friendly-prefixed placeholder for `OTHERSECRET` before the regex // value it collides with (`tok_abc123`, sanitized to `TOKABC123`, matching the // `friendlyName`) is even discovered in a LATER message — baking a normalized // rendering of that later secret into provider-visible text as an "innocent" // friendly label instead of stripping it to a bare placeholder. const obfuscator = new SecretObfuscator([ { type: "plain", content: "OTHERSECRET", friendlyName: "TOKABC123" }, { type: "regex", content: "tok_[a-z0-9]+" }, ]); const messages: Message[] = [ { role: "user", content: "first message carries OTHERSECRET", timestamp: 1 }, { role: "user", content: "later message reveals tok_abc123", timestamp: 2 }, ]; const obfuscated = obfuscateMessages(obfuscator, messages); const serialized = JSON.stringify(obfuscated); expect(serialized).not.toContain("OTHERSECRET"); expect(serialized).not.toContain("tok_abc123"); // The friendly prefix is itself a normalized rendering of the later-discovered // regex value; sharing regex values across the batch up front must strip it // down to a bare placeholder rather than bake it into message 1's output. expect(serialized).not.toContain("TOKABC123_"); // Both originals still round-trip through deobfuscation of the serialized batch. const restored = obfuscator.deobfuscate(serialized); expect(restored).toContain("OTHERSECRET"); expect(restored).toContain("tok_abc123"); // Re-obfuscating the already-obfuscated batch is a fixed point: identical bytes, // so re-running obfuscateMessages over a growing conversation never busts the // provider prompt cache for messages already sent. expect(JSON.stringify(obfuscateMessages(obfuscator, obfuscated))).toEqual(serialized); }); it("includes assistant replay content in the batch regex collision pre-scan", () => { const obfuscator = new SecretObfuscator([ { type: "plain", content: "OTHERSECRET", friendlyName: "TOKABC123" }, { type: "regex", content: "tok_[a-z0-9]+" }, ]); const messages: Message[] = [ { role: "user", content: "first message carries OTHERSECRET", timestamp: 1 }, { role: "assistant", content: [{ type: "toolCall", id: "call-1", name: "read", arguments: { token: "tok_abc123" } }], api: "anthropic-messages", provider: "anthropic", model: "test-model", usage: { input: 1, output: 1, cacheRead: 0, cacheWrite: 0, totalTokens: 2, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 }, }, stopReason: "toolUse", timestamp: 2, }, ]; const serialized = JSON.stringify(obfuscateMessages(obfuscator, messages)); expect(serialized).not.toContain("OTHERSECRET"); expect(serialized).not.toContain("tok_abc123"); expect(serialized).not.toContain("TOKABC123_"); }); it("collects regex-protected values across the whole tool-call argument payload before obfuscating any single string", () => { // Regression: `obfuscateToolArguments` must precompute regex-protected // values by walking the ENTIRE JSON argument object up front — the same // whole-batch precomputation `obfuscateMessages` performs above — before // redacting any single string within it, when the caller passes no // `sharedRegexSecretValues` of its own. Mapping each JSON string // independently (obfuscating `first` before `nested.later` is ever // visited) would let `first`'s OTHERSECRET mint a friendly-prefixed // placeholder before `tok_abc123` is discovered deeper in the same // payload, baking a normalized rendering of that later secret into // provider-visible text as an "innocent" friendly label. const obfuscator = new SecretObfuscator([ { type: "plain", content: "OTHERSECRET", friendlyName: "TOKABC123" }, { type: "regex", content: "tok_[a-z0-9]+" }, ]); const args = { first: "OTHERSECRET", nested: { later: "tok_abc123" } }; const obfuscated = obfuscateToolArguments(obfuscator, args); const serialized = JSON.stringify(obfuscated); expect(serialized).not.toContain("OTHERSECRET"); expect(serialized).not.toContain("tok_abc123"); // The friendly prefix is itself a normalized rendering of the // later-discovered regex value; collecting regex values across the whole // payload up front must strip it down to a bare placeholder rather than // bake it into `first`'s output before `nested.later` is ever visited. expect(serialized).not.toContain("TOKABC123_"); // Both originals round-trip through deobfuscation of the serialized args. expect(deobfuscateToolArguments(obfuscator, obfuscated)).toEqual(args); // Re-obfuscating the already-obfuscated args is a fixed point: identical // bytes, matching the cache-stability guarantee `obfuscateMessages` gives // full conversation batches. expect(JSON.stringify(obfuscateToolArguments(obfuscator, obfuscated))).toEqual(serialized); }); it("strips a stale friendly prefix from persisted assistant history once a later message reveals the colliding regex secret", () => { // Regression: `obfuscateAssistantContentForReplay` only scans // the assistant content passed to THIS call. If an earlier turn already // minted `#TOKABC123_#` for `OTHERSECRET` and that exact placeholder // was persisted verbatim into assistant history before `tok_abc123` was // ever seen, replaying that history alongside a NEW message that finally // reveals `tok_abc123` must still scrub the now-unsafe prefix out of the // OLD assistant text — not merely avoid minting a fresh prefixed one. const obfuscator = new SecretObfuscator([ { type: "plain", content: "OTHERSECRET", friendlyName: "TOKABC123" }, { type: "regex", content: "tok_[a-z0-9]+" }, ]); function assistantText(message: Message): string { if (message.role !== "assistant") throw new Error("expected an assistant message"); const block = message.content.find((c): c is TextContent => c.type === "text"); if (block === undefined) throw new Error("expected an assistant text block"); return block.text; } function toolResultText(message: Message): string { if (message.role !== "toolResult") throw new Error("expected a toolResult message"); const block = message.content.find((c): c is TextContent => c.type === "text"); if (block === undefined) throw new Error("expected a toolResult text block"); return block.text; } function userText(message: Message): string { if (message.role !== "user") throw new Error("expected a user message"); if (typeof message.content !== "string") throw new Error("expected string user content"); return message.content; } // Turn 1: mint the friendly-prefixed placeholder for OTHERSECRET while // tok_abc123 is still unknown to this obfuscator instance — nothing in // this batch matches the regex pattern, so the collision guard has // nothing to compare the friendly name against yet. const minted = obfuscateMessages(obfuscator, [ { role: "user", content: "remember OTHERSECRET for later", timestamp: 1 }, ]); const mintedMatch = /\$\$TOKABC123_[A-Z0-9]+(?::[ULCM])?\$\$/.exec(userText(minted[0])); if (mintedMatch === null) throw new Error("expected turn 1 to mint a friendly-prefixed placeholder"); const mintedPlaceholder = mintedMatch[0]; const bareAlias = mintedPlaceholder.replace(/^\$\$TOKABC123_/, "$$"); // Turn 2: assistant history persisted that exact prefixed placeholder // verbatim, wrapped in ordinary prose that carries no secret material // and must survive untouched. A tool result in the same batch finally // reveals the raw regex-protected secret it collides with. const untouchedPrefix = "Sure, I stored it — "; const untouchedSuffix = " — anything else?"; const assistantMessage: Message = { role: "assistant", content: [{ type: "text", text: `${untouchedPrefix}${mintedPlaceholder}${untouchedSuffix}` }], api: "anthropic-messages", provider: "anthropic", model: "test-model", usage: { input: 1, output: 1, cacheRead: 0, cacheWrite: 0, totalTokens: 2, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 }, }, stopReason: "stop", timestamp: 2, }; const toolResultMessage: Message = { role: "toolResult", toolCallId: "call-1", toolName: "bash", content: [{ type: "text", text: "the real token is tok_abc123" }], isError: false, timestamp: 3, }; const result = obfuscateMessages(obfuscator, [assistantMessage, toolResultMessage]); const strippedAssistantText = assistantText(result[0]); const redactedToolResultText = toolResultText(result[1]); // The stale friendly prefix is gone, replaced by the friendly-name- // independent bare alias minted back in turn 1... expect(strippedAssistantText).not.toContain("TOKABC123_"); expect(strippedAssistantText).toContain(bareAlias); // ...while the surrounding raw assistant prose is untouched byte-for-byte. expect(strippedAssistantText).toContain(untouchedPrefix); expect(strippedAssistantText).toContain(untouchedSuffix); // The newly revealed raw secret is redacted in the tool result. expect(redactedToolResultText).not.toContain("tok_abc123"); // Both secrets still round-trip through deobfuscation of the serialized batch. const serialized = JSON.stringify(result); const restored = obfuscator.deobfuscate(serialized); expect(restored).toContain("OTHERSECRET"); expect(restored).toContain("tok_abc123"); // Fixed point: re-obfuscating the already-stripped batch changes nothing further. expect(JSON.stringify(obfuscateMessages(obfuscator, result))).toEqual(serialized); }); it("strips a stale friendly prefix from an assistant toolCall block's arguments, intent, and rawBlock once a later message reveals the colliding regex secret", () => { // Regression: `obfuscateAssistantContentForReplay` walks `text` // blocks and `toolCall` blocks differently — a toolCall's `arguments`, // `intent`, and `rawBlock` are where the model's own tool invocations // persist a friendly-prefixed placeholder minted in an earlier turn. If a // later message reveals the regex-protected value that placeholder's // friendly name normalizes to, the toolCall fields must be re-scanned and // scrubbed exactly like assistant text — leaving "TOKABC123_" baked into a // replayed tool call would expose the very value `tok_[a-z0-9]+` is // configured to hide, just via a different content-block type. const obfuscator = new SecretObfuscator([ { type: "plain", content: "OTHERSECRET", friendlyName: "TOKABC123" }, { type: "regex", content: "tok_[a-z0-9]+" }, ]); type AssistantToolCallBlock = Extract; function assistantToolCall(message: Message): AssistantToolCallBlock { if (message.role !== "assistant") throw new Error("expected an assistant message"); const block = message.content.find((c): c is AssistantToolCallBlock => c.type === "toolCall"); if (block === undefined) throw new Error("expected an assistant toolCall block"); return block; } function toolResultText(message: Message): string { if (message.role !== "toolResult") throw new Error("expected a toolResult message"); const block = message.content.find((c): c is TextContent => c.type === "text"); if (block === undefined) throw new Error("expected a toolResult text block"); return block.text; } function userText(message: Message): string { if (message.role !== "user") throw new Error("expected a user message"); if (typeof message.content !== "string") throw new Error("expected string user content"); return message.content; } // Turn 1: mint the friendly-prefixed placeholder for OTHERSECRET while // tok_abc123 is still unknown to this obfuscator instance. const minted = obfuscateMessages(obfuscator, [ { role: "user", content: "remember OTHERSECRET for later", timestamp: 1 }, ]); const mintedMatch = /\$\$TOKABC123_[A-Z0-9]+(?::[ULCM])?\$\$/.exec(userText(minted[0])); if (mintedMatch === null) throw new Error("expected turn 1 to mint a friendly-prefixed placeholder"); const mintedPlaceholder = mintedMatch[0]; const bareAlias = mintedPlaceholder.replace(/^\$\$TOKABC123_/, "$$"); // Turn 2: assistant history persisted that exact prefixed placeholder // verbatim inside a toolCall block's arguments, intent, and rawBlock — // each also carrying ordinary raw prose that carries no secret material // and must survive untouched. A tool result in the same batch finally // reveals the raw regex-protected secret it collides with. const untouchedPrefix = "run with token "; const untouchedSuffix = " please"; const staleField = `${untouchedPrefix}${mintedPlaceholder}${untouchedSuffix}`; const assistantMessage: Message = { role: "assistant", content: [ { type: "toolCall", id: "call-1", name: "bash", arguments: { command: staleField, note: "no secret material here" }, intent: staleField, rawBlock: staleField, }, ], api: "anthropic-messages", provider: "anthropic", model: "test-model", usage: { input: 1, output: 1, cacheRead: 0, cacheWrite: 0, totalTokens: 2, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 }, }, stopReason: "stop", timestamp: 2, }; const toolResultMessage: Message = { role: "toolResult", toolCallId: "call-1", toolName: "bash", content: [{ type: "text", text: "the real token is tok_abc123" }], isError: false, timestamp: 3, }; const result = obfuscateMessages(obfuscator, [assistantMessage, toolResultMessage]); const strippedToolCall = assistantToolCall(result[0]); const redactedToolResultText = toolResultText(result[1]); // The stale friendly prefix is gone from every toolCall field, replaced by // the friendly-name-independent bare alias minted back in turn 1, while the // surrounding raw prose in each field is untouched byte-for-byte. const strippedArgsJson = JSON.stringify(strippedToolCall.arguments); expect(strippedArgsJson).not.toContain("TOKABC123_"); expect(strippedArgsJson).toContain(bareAlias); expect(strippedArgsJson).toContain(untouchedPrefix); expect(strippedArgsJson).toContain(untouchedSuffix); if (strippedToolCall.intent === undefined) throw new Error("expected intent to survive stripping"); expect(strippedToolCall.intent).not.toContain("TOKABC123_"); expect(strippedToolCall.intent).toContain(bareAlias); expect(strippedToolCall.intent).toContain(untouchedPrefix); expect(strippedToolCall.intent).toContain(untouchedSuffix); if (strippedToolCall.rawBlock === undefined) throw new Error("expected rawBlock to survive stripping"); expect(strippedToolCall.rawBlock).not.toContain("TOKABC123_"); expect(strippedToolCall.rawBlock).toContain(bareAlias); expect(strippedToolCall.rawBlock).toContain(untouchedPrefix); expect(strippedToolCall.rawBlock).toContain(untouchedSuffix); // The newly revealed raw secret is redacted in the tool result. expect(redactedToolResultText).not.toContain("tok_abc123"); // Both secrets still round-trip through deobfuscation of the serialized batch. const serialized = JSON.stringify(result); const restored = obfuscator.deobfuscate(serialized); expect(restored).toContain("OTHERSECRET"); expect(restored).toContain("tok_abc123"); // Fixed point: re-obfuscating the already-stripped batch changes nothing further. expect(JSON.stringify(obfuscateMessages(obfuscator, result))).toEqual(serialized); }); it("strips a stale friendly prefix from an assistant thinking block once a later message reveals the colliding regex secret", () => { // Regression: `obfuscateAssistantContentForReplay` strips // `text` and `toolCall` blocks, but everywhere ELSE in the obfuscator a // `thinking` block is deliberately treated as opaque provider-replay data // that passes through byte-identical (see `deobfuscateAssistantContent`'s // doc comment). If the friendly-prefix strip pass fell through thinking // blocks the same way, a stale `#TOKABC123_#` placeholder minted // before `tok_abc123` was ever seen would keep leaking the friendly // name's normalized collision with that regex-protected secret every // time persisted thinking history was replayed to the provider. const obfuscator = new SecretObfuscator([ { type: "plain", content: "OTHERSECRET", friendlyName: "TOKABC123" }, { type: "regex", content: "tok_[a-z0-9]+" }, ]); type AssistantThinkingBlock = Extract; function assistantThinking(message: Message): AssistantThinkingBlock { if (message.role !== "assistant") throw new Error("expected an assistant message"); const block = message.content.find((c): c is AssistantThinkingBlock => c.type === "thinking"); if (block === undefined) throw new Error("expected an assistant thinking block"); return block; } function toolResultText(message: Message): string { if (message.role !== "toolResult") throw new Error("expected a toolResult message"); const block = message.content.find((c): c is TextContent => c.type === "text"); if (block === undefined) throw new Error("expected a toolResult text block"); return block.text; } function userText(message: Message): string { if (message.role !== "user") throw new Error("expected a user message"); if (typeof message.content !== "string") throw new Error("expected string user content"); return message.content; } // Turn 1: mint the friendly-prefixed placeholder for OTHERSECRET while // tok_abc123 is still unknown to this obfuscator instance. const minted = obfuscateMessages(obfuscator, [ { role: "user", content: "remember OTHERSECRET for later", timestamp: 1 }, ]); const mintedMatch = /\$\$TOKABC123_[A-Z0-9]+(?::[ULCM])?\$\$/.exec(userText(minted[0])); if (mintedMatch === null) throw new Error("expected turn 1 to mint a friendly-prefixed placeholder"); const mintedPlaceholder = mintedMatch[0]; const bareAlias = mintedPlaceholder.replace(/^\$\$TOKABC123_/, "$$"); // Turn 2: assistant history persisted that exact prefixed placeholder // verbatim inside a thinking block, wrapped in ordinary raw reasoning // prose that carries no secret material and must survive untouched. A // tool result in the same batch finally reveals the raw regex-protected // secret it collides with. const untouchedPrefix = "Let me recall the value I stored earlier: "; const untouchedSuffix = " — that should still be correct."; const assistantMessage: Message = { role: "assistant", content: [ { type: "thinking", thinking: `${untouchedPrefix}${mintedPlaceholder}${untouchedSuffix}`, thinkingSignature: "signed-original-thinking", }, ], api: "anthropic-messages", provider: "anthropic", model: "test-model", usage: { input: 1, output: 1, cacheRead: 0, cacheWrite: 0, totalTokens: 2, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 }, }, stopReason: "stop", timestamp: 2, }; const toolResultMessage: Message = { role: "toolResult", toolCallId: "call-1", toolName: "bash", content: [{ type: "text", text: "the real token is tok_abc123" }], isError: false, timestamp: 3, }; const result = obfuscateMessages(obfuscator, [assistantMessage, toolResultMessage]); const strippedThinkingBlock = assistantThinking(result[0]); const strippedThinking = strippedThinkingBlock.thinking; const redactedToolResultText = toolResultText(result[1]); // The stale friendly prefix is gone from the thinking block, replaced by // the friendly-name-independent bare alias minted back in turn 1, while // the surrounding raw reasoning prose is untouched byte-for-byte. expect(strippedThinking).not.toContain("TOKABC123_"); expect(strippedThinking).toContain(bareAlias); expect(strippedThinking).toContain(untouchedPrefix); expect(strippedThinking).toContain(untouchedSuffix); expect(strippedThinkingBlock.thinkingSignature).toBeUndefined(); // The newly revealed raw secret is redacted in the tool result. expect(redactedToolResultText).not.toContain("tok_abc123"); // Both secrets still round-trip through deobfuscation of the serialized batch. const serialized = JSON.stringify(result); const restored = obfuscator.deobfuscate(serialized); expect(restored).toContain("OTHERSECRET"); expect(restored).toContain("tok_abc123"); // Fixed point: re-obfuscating the already-stripped batch changes nothing further. expect(JSON.stringify(obfuscateMessages(obfuscator, result))).toEqual(serialized); }); }); describe("deobfuscateAgentMessages (display restore)", () => { it("restores assistant text and tool calls while leaving raw user text and thinking untouched", () => { const secret = "DISPLAY_SECRET_TOKEN_123"; const obfuscator = new SecretObfuscator([{ type: "plain", content: secret }]); const placeholder = obfuscator.obfuscate(secret); expect(placeholder).not.toBe(secret); const userMsg: AgentMessage = { role: "user", content: `literal ${placeholder} token`, timestamp: 1 }; const assistantMsg: AgentMessage = { role: "assistant", content: [ { type: "text", text: `answer ${placeholder}` }, { type: "thinking", thinking: `reason ${placeholder}` }, { type: "toolCall", id: "call_1", name: "read", arguments: { path: `path ${placeholder}` }, intent: `intent ${placeholder}`, }, ], api: "test", provider: "test", model: "test", usage: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, totalTokens: 0, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 }, }, stopReason: "toolUse", timestamp: 2, }; const branchSummary: AgentMessage = { role: "branchSummary", summary: `branch ${placeholder}`, fromId: "x", timestamp: 3, }; const compactionSummary: AgentMessage = { role: "compactionSummary", summary: `compact ${placeholder}`, shortSummary: `short ${placeholder}`, tokensBefore: 0, timestamp: 4, }; const restored = deobfuscateAgentMessages(obfuscator, [userMsg, assistantMsg, branchSummary, compactionSummary]); // Assistant text and tool-call args/intent are restored to the real secret. const restoredAssistant = restored[1] as AssistantMessage; const assistantJson = JSON.stringify(restoredAssistant.content); expect(assistantJson).toContain(secret); expect(assistantJson).not.toContain(`answer ${placeholder}`); expect(assistantJson).not.toContain(`path ${placeholder}`); expect(assistantJson).not.toContain(`intent ${placeholder}`); // Opaque thinking is never walked: placeholder-shaped bytes survive unchanged. expect(assistantJson).toContain(`reason ${placeholder}`); expect(assistantJson).not.toContain(`reason ${secret}`); // Model-generated summaries are restored. expect((restored[2] as { summary: string }).summary).toBe(`branch ${secret}`); expect((restored[3] as { summary: string; shortSummary?: string }).summary).toBe(`compact ${secret}`); expect((restored[3] as { summary: string; shortSummary?: string }).shortSummary).toBe(`short ${secret}`); // The user message is persisted raw and never walked: a literal placeholder-shaped token // survives byte-identical (same reference) rather than being turned into the secret. expect(restored[0]).toBe(userMsg); }); it("restores compactionSummary block text while leaving snapcompact image bytes intact", () => { const secret = "BLOCKS_SECRET_TOKEN_456"; const obfuscator = new SecretObfuscator([{ type: "plain", content: secret }]); const placeholder = obfuscator.obfuscate(secret); const imageData = `frame${secret}bytes==`; const message: AgentMessage = { role: "compactionSummary", summary: `summary ${placeholder}`, tokensBefore: 0, blocks: [ { type: "text", text: `archived ${placeholder}` }, { type: "image", data: imageData, mimeType: "image/png" }, ], timestamp: 1, }; const restored = deobfuscateAgentMessages(obfuscator, [message])[0]; if (restored?.role !== "compactionSummary") throw new Error("expected restored compaction summary"); const blocks = restored.blocks ?? []; const text = blocks[0]; const image = blocks[1]; // Archived text is restored to the real secret... expect(text.type === "text" && text.text).toBe(`archived ${secret}`); // ...while the snapcompact image bytes pass through untouched. expect(image.type === "image" && image.data).toBe(imageData); }); });