smashimo/oh-my-pi
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(coding-agent/session): prevented data loss in session serialization

Browse Source 
- Persist signed message blocks (`text`, `thinking`, `toolCall`) and encrypted reasoning payloads verbatim during session serialization instead of clearing or truncating them.
- Preserve signature keys instead of replacing them with empty strings when they exceed persistence size limits.
- Exempt official first-party OpenAI and Anthropic API endpoints from the leaked-thinking stream healing wrapper to prevent misfires on legitimate visible text fences.
This commit is contained in:
can1357
2026-07-02 03:57:35 +02:00
parent 0d96c2b6ec
commit c4c0331345
17 changed files with 158 additions and 152 deletions
Download Patch File Download Diff File Expand all files Collapse all files
packages/agent/CHANGELOG.md
+1 -2
View File
@@ -8,8 +8,7 @@
### Fixed
- Fixed an issue where legacy steering messages were prematurely consumed and dropped during in-flight tool execution polls
- Fixed an issue where legacy steering messages were prematurely consumed and dropped during in-flight tool execution polls.
- Fixed an issue where skipped tool results in queued messages were incorrectly treated as completed, preventing necessary retries.
- Improved branch summaries to preserve informative tool results from abandoned branches while filtering out redundant output.
- Fixed interruptible tool waits to properly abort on host-provided IRC interrupts in addition to user steering.
packages/ai/CHANGELOG.md
+8 -9
View File
@@ -4,21 +4,20 @@
### Added
- Added opt-in support for Anthropic's server-side fallback beta (`server-side-fallback-2026-06-01`) on the `anthropic-messages` provider, including support for `AnthropicOptions.fallbacks`, mid-stream fallback content blocks, fallback billing/usage iterations, and automatic filtering of fallback blocks during cross-provider message transformations.
- Added opt-in support for Anthropic's server-side fallback beta (server-side-fallback-2026-06-01) on the anthropic-messages provider, including support for AnthropicOptions.fallbacks and automatic filtering of fallback blocks during cross-provider message transformations.
### Changed
- Skip leaked-thinking stream healing for official first-party Anthropic, OpenAI, and Codex endpoints
- Updated CoreWeave Serverless Inference login instructions to clarify persisting `COREWEAVE_PROJECT` in shell startup files.
- Improved stream healing for official first-party endpoints (Anthropic, OpenAI, and OpenAI Codex) by skipping leaked-thinking healing, preventing misfires on legitimate code blocks while maintaining healing for third-party gateways and custom base URLs.
- Updated CoreWeave Serverless Inference login instructions to clarify persisting COREWEAVE_PROJECT in shell startup files.
### Fixed
- Fixed an issue where broker usage fetch failures were not cached, causing redundant network requests during sequential ranking passes when the broker is offline.
- Fixed Xiaomi MiMo API key validation to use the supported `mimo-v2.5` model.
- Fixed certificate verification errors for custom gateways behind private CA bundles by ensuring `NODE_EXTRA_CA_CERTS` is respected across all provider fetches (including OpenAI-compatible, Codex, Ollama, Azure, and Google).
- Fixed Claude Fable demoted-thinking replay to use markdown-italic assistant prose instead of `<thinking>` tags, preventing context issues after model switches.
- Fixed OpenAI Responses replay errors (400 Bad Request) caused by missing reasoning items in locally rebuilt assistant item IDs during history replay.
- Fixed a performance issue where broker usage fetch failures were not cached, causing redundant network requests when the broker is offline.
- Fixed Xiaomi MiMo API key validation to use the supported mimo-v2.5 model.
- Fixed certificate verification errors for custom gateways behind private CA bundles by ensuring NODE_EXTRA_CA_CERTS is respected across all provider fetches.
- Fixed Claude Fable demoted-thinking replay to use markdown-italic assistant prose instead of <thinking> tags, preventing context issues after model switches.
- Fixed OpenAI Responses replay errors (400 Bad Request) caused by missing reasoning items during history replay.
## [16.2.13] - 2026-07-01
packages/ai/src/stream.ts
+2 -1
View File
@@ -99,7 +99,8 @@ function isGoogleVertexAuthenticatedModel(model: Model<Api>): boolean {
* gateway may well leak. URL checks are strict (exact origin / path boundary
* or parsed hostname) — a substring match would accept lookalikes like
* `https://api.openai.com.evil/`. Anthropic Foundry (`CLAUDE_CODE_USE_FOUNDRY`)
* dispatches an empty `baseUrl` to an enterprise gateway, so it is never exempt.
* redirects an empty `baseUrl` to `FOUNDRY_BASE_URL`, so the check runs against
* that effective endpoint — exempt only when it resolves to the official host.
*/
function isLeakedThinkingHealExempt(model: Model<Api>): boolean {
switch (model.provider) {
packages/ai/src/utils/leaked-thinking-stream.ts
+8 -4
View File
@@ -3,11 +3,15 @@
*
* Some providers emit their canonical reasoning idioms (` ```thinking `,
* `<think>`, Gemma/Harmony channels, …) into the *visible* text stream instead
* of a structured thinking part. {@link wrapLeakedThinkingStream} re-projects any
* of a structured thinking part. {@link wrapLeakedThinkingStream} re-projects a
* provider stream into a fresh {@link AssistantMessageEventStream}, splitting the
* leaked fences out into proper `thinking` blocks *live* as deltas arrive — so
* every provider gets the same healing, not just the three with provider-local
* {@link StreamMarkupHealing} loops.
* leaked fences out into proper `thinking` blocks *live* as deltas arrive.
*
* Applied to every provider stream *except* official first-party endpoints
* (the official Anthropic API and the official OpenAI / OpenAI-Codex endpoints),
* which return structured thinking and never leak — `healLeakedThinking` in
* `../stream.ts` gates the wrap so the healer cannot misfire on legitimate
* fenced content those models emit as visible text.
*
* The healing is idempotent: a second pass over already-clean text finds no
* fences, so wrapping a provider that already heals (or wrapping twice) is a
packages/ai/test/leaked-thinking-stream.test.ts
+8 -4
View File
@@ -446,10 +446,14 @@ describe("leaked thinking healing through stream()", () => {
it("splits a leaked fence for a non-official anthropic-messages endpoint", async () => {
// A third-party gateway reusing the anthropic-messages wire format may leak,
// so the central wrapper still heals when the endpoint is not official.
const result = await stream(anthropicModel({ provider: "z-ai", baseUrl: "https://api.z.ai/api/anthropic" }), context, {
apiKey: "test",
fetch: anthropicLeakFetch(leaked),
}).result();
const result = await stream(
anthropicModel({ provider: "zai", baseUrl: "https://api.z.ai/api/anthropic" }),
context,
{
apiKey: "test",
fetch: anthropicLeakFetch(leaked),
},
).result();
expect(result.content.map(b => b.type)).toEqual(["thinking", "text"]);
const thinking = thinks(result)
packages/ai/test/stream-markup-healing.test.ts
+29
View File
@@ -246,6 +246,35 @@ describe("openai-completions leaked thinking healing", () => {
});
});
describe("official OpenAI leaked thinking healing exemption", () => {
// The official OpenAI endpoint returns structured reasoning and never leaks
// fences, so neither the provider-local healer nor the central
// wrapLeakedThinkingStream wrap runs — a ` ```thinking ` block the model chose
// to write must stay verbatim visible text. Routed through stream() (not the
// provider directly) so both gates are exercised.
const officialOpenAI = getBundledModel("openai", "gpt-5.5");
const completionsModel = buildModel({
...officialOpenAI,
api: "openai-completions",
});
it("leaves a leaked fence intact for the official OpenAI endpoint", async () => {
const leaked = "```thinking\nWeigh the options.\n```\nFinal answer.";
const result = await stream(completionsModel, baseContext(), {
apiKey: "test",
fetch: mockFetch([
chunk(completionsModel.id, { content: leaked }),
chunk(completionsModel.id, {}, "stop"),
"[DONE]",
]),
}).result();
expect(result.content.map(b => b.type)).toEqual(["text"]);
expect(result.content.filter((b): b is ThinkingContent => b.type === "thinking")).toHaveLength(0);
expect(result.content.map(b => (b.type === "text" ? b.text : "")).join("")).toBe(leaked);
});
});
describe("google-gemini-cli leaked thinking healing", () => {
it("lifts a leaked Gemini thinking fence before a native tool call", async () => {
const model = geminiCliModel();
packages/catalog/CHANGELOG.md
+4 -3
View File
@@ -8,10 +8,11 @@
### Fixed
- Fixed the Xiaomi provider's default model to use the supported `mimo-v2.5` model.
- Fixed model discovery probes (including Ollama and metadata fetches) failing behind private-CA gateways by ensuring they honor `NODE_EXTRA_CA_CERTS`.
- Fixed stream markup healing pattern misfires by disabling the healer on the official OpenAI endpoint.
- Updated the Xiaomi provider's default model to the supported `mimo-v2.5` model.
- Fixed model discovery probes (including Ollama and metadata fetches) failing behind private-CA gateways by ensuring they honor the `NODE_EXTRA_CA_CERTS` environment variable.
- Fixed CoreWeave Serverless Inference project-header detection to ensure blank OpenAI-Project overrides do not block the `COREWEAVE_PROJECT` fallback.
- Fixed LiteLLM MiniMax M3 discovery to remove reseller-only display suffixes, and invalidated the model cache to ensure stale suffixes are cleared immediately.
- Fixed LiteLLM MiniMax M3 discovery to remove reseller-only display suffixes and invalidated the model cache to clear stale suffixes immediately.
## [16.2.13] - 2026-07-01
packages/catalog/src/compat/openai.ts
+25 -6
View File
@@ -117,20 +117,39 @@ function resolveReasoningDisableMode(
/**
* Pick the leaked-markup healer for an OpenAI-compatible visible-text stream.
* Kimi chat-template tokens and DeepSeek DSML envelopes need their dedicated
* tool-call grammars; every other model defaults to `"thinking"`. All patterns
* run the generic thinking healer, so leaked reasoning idioms (e.g. a Gemini
* ` ```thinking ` fence on OpenRouter) are always recovered from `delta.content`.
* tool-call grammars. Every other OpenAI-compatible model defaults to
* `"thinking"` so leaked reasoning idioms (e.g. a Gemini ` ```thinking ` fence
* on OpenRouter) are recovered from `delta.content` — **except** the official
* OpenAI endpoint (`provider: "openai"` + `api.openai.com`), which returns
* structured reasoning and never leaks, so it heals nothing (returns
* `undefined`) to avoid misfiring on legitimate fenced content.
*/
function detectStreamMarkupHealingPattern(provider: string, modelId: string): OpenAIStreamMarkupHealingPattern {
function detectStreamMarkupHealingPattern(
provider: string,
modelId: string,
baseUrl: string,
): OpenAIStreamMarkupHealingPattern | undefined {
if (provider === "kimi-code" || provider === "moonshot" || /kimi[-/_.]?k2/i.test(modelId)) {
return "kimi";
}
if (isDeepseekModelIdOrName(modelId) && DSML_HEALING_PROVIDERS.has(provider)) {
return "dsml";
}
if (isOfficialOpenAIEndpoint(provider, baseUrl)) return undefined;
return "thinking";
}
/** Strict official-OpenAI check: provider id `openai` and an `api.openai.com` host (missing baseUrl defaults there). */
function isOfficialOpenAIEndpoint(provider: string, baseUrl: string): boolean {
if (provider !== "openai") return false;
if (!baseUrl) return true;
try {
return new URL(baseUrl).hostname === "api.openai.com";
} catch {
return false;
}
}
/**
* OpenCode's gateways (https://opencode.ai/zen|go) gate `reasoning_content`
* on the request's thinking state for every model they front (Kimi K2.x,
@@ -517,7 +536,7 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
streamIdleTimeoutMs,
stripDeepseekSpecialTokens:
isDeepseekModelIdOrName(spec.id) && (provider === "nvidia" || provider === "deepseek"),
streamMarkupHealingPattern: detectStreamMarkupHealingPattern(provider, spec.id),
streamMarkupHealingPattern: detectStreamMarkupHealingPattern(provider, spec.id, baseUrl),
reasoningDeltasMayBeCumulative:
MINIMAX_PROVIDER_OR_ID_PATTERN.test(provider) || MINIMAX_PROVIDER_OR_ID_PATTERN.test(spec.id),
emptyLengthFinishIsContextError: provider === "ollama",
@@ -646,7 +665,7 @@ export function buildOpenAIResponsesCompat(spec: OpenAIResponsesSpecLike): Resol
supportsObfuscationOptOut: isOpenAIUrl || spec.provider === "openai",
stripDeepseekSpecialTokens:
Boolean(id) && isDeepseekModelIdOrName(id) && (spec.provider === "nvidia" || spec.provider === "deepseek"),
streamMarkupHealingPattern: id ? detectStreamMarkupHealingPattern(spec.provider, id) : undefined,
streamMarkupHealingPattern: id ? detectStreamMarkupHealingPattern(spec.provider, id, baseUrl) : undefined,
reasoningDeltasMayBeCumulative:
MINIMAX_PROVIDER_OR_ID_PATTERN.test(spec.provider) || (id ? MINIMAX_PROVIDER_OR_ID_PATTERN.test(id) : false),
emptyLengthFinishIsContextError: spec.provider === "ollama",
packages/catalog/test/build.test.ts
+24
View File
@@ -297,6 +297,30 @@ describe("openai-completions wire-quirk compat detection", () => {
expect(buildOpenAICompat(completionsSpec()).dropThinkingWhenReasoningEffort).toBe(false);
});
it("disables the leaked-markup healer for the official OpenAI endpoint only", () => {
// Official OpenAI returns structured reasoning and never leaks fences, so
// the provider-local healer stays off; every other OpenAI-compatible host
// keeps the default "thinking" healer, and Kimi/DSML keep their grammars.
expect(
buildOpenAICompat(completionsSpec({ provider: "openai", baseUrl: "https://api.openai.com/v1" }))
.streamMarkupHealingPattern,
).toBeUndefined();
expect(
buildOpenAICompat(completionsSpec({ provider: "openrouter", baseUrl: "https://openrouter.ai/api/v1" }))
.streamMarkupHealingPattern,
).toBe("thinking");
// A lookalike host under the openai provider id is NOT the official endpoint.
expect(
buildOpenAICompat(completionsSpec({ provider: "openai", baseUrl: "https://api.openai.com.evil/v1" }))
.streamMarkupHealingPattern,
).toBe("thinking");
expect(
buildOpenAICompat(
completionsSpec({ provider: "moonshot", id: "kimi-k2", baseUrl: "https://api.moonshot.ai/v1" }),
).streamMarkupHealingPattern,
).toBe("kimi");
});
it("derives Responses obfuscation opt-out and wire mode per surface", () => {
expect(
buildOpenAIResponsesCompat({
packages/coding-agent/CHANGELOG.md
+12 -62
View File
@@ -4,7 +4,7 @@
### Added
- Added `providers.anthropic.serverSideFallback` configuration option to opt into Anthropic's server-side-fallback beta chain, allowing Claude Fable 5 / Mythos 5 requests to automatically retry on Opus 4.8 when blocked by classifiers.
- Added `providers.anthropic.serverSideFallback` configuration option to opt into Anthropic's server-side-fallback beta chain, allowing Claude requests to automatically retry on alternative models when blocked by classifiers.
- Added `task.softRequestBudgetNotice` configuration option to enable subagent soft-budget wrap-up steering notices while keeping the graceful abort guard active.
### Changed
@@ -16,66 +16,7 @@
### Fixed
- Fixed subagent HUD layout and tree rendering in the TUI to align correctly with other HUD panels
- Fixed session persistence corrupting Anthropic signed thinking blocks: the generic 500K-character truncation cap could shorten a large `thinking` string while leaving its cryptographic signature intact, so replaying or resuming the session sent a signature that no longer matched the text and the provider rejected it with an HTTP 400. Signed `thinking` blocks and encrypted `redactedThinking` blobs are now persisted verbatim (all-or-nothing); unsigned thinking and plain text remain truncatable for size control.
- Fixed post-rewind context to tell the agent the checkpoint completed and to make repeat `rewind` calls recover with guidance instead of a bare no-checkpoint error; the branch-scan rehydration now also restores an active checkpoint when a session resumes with the latest checkpoint not yet rewound, so `rewind` can complete instead of failing with "No active checkpoint". ([#4187](https://github.com/can1357/oh-my-pi/issues/4187))
- Fixed task.maxConcurrency being breachable when a queued spawn was cancelled: the spawn path could release a semaphore permit it never acquired, letting a later task start while the cap was saturated.
- Fixed session exit diagnostics recording signal and crash exits (SIGTERM, SIGHUP, uncaught exceptions) as a normal "dispose": the postmortem teardown now threads the real reason into session disposal.
- Fixed the subagent yield-label guard ignoring JTD discriminator (oneOf) output schemas, which let stale incremental labels pass into successful results when final validation was skipped after retries.
- Fixed grep/ast_grep search scopes rejecting `www.` and collapsed-scheme (`https:/host`) URL spellings that the read tool accepts; unresolvable URL-shaped scopes now fail with an explicit external-URL error instead of "Path not found".
- Fixed model discovery ignoring `NODE_EXTRA_CA_CERTS`: the model registry's default fetch now applies the extra-CA wrapper, so `/models` probes work behind private-CA gateways like provider chat requests.
- Fixed ctrl+p role-model cycling getting stuck on one transition and skipping every other role: a session-branch traversal regression returned entries leaf-to-root, so the cycle (and session model restore) read the oldest recorded model change instead of the newest.
- Fixed ctrl+p cycling from a stale slot after the model was switched through another surface (alt+m, /model, retry fallback): the recorded role is now trusted only while its resolved model is still the active model, falling back to matching by model.
- Fixed the apply_patch tool to prevent silently overwriting pre-existing files during creation or renaming, rejecting upfront with an error instead.
- Fixed multi-file apply_patch to stop at the first failing file, surface applied vs. skipped paths, and correctly report the error to the agent loop.
- Fixed process termination (SIGTERM, SIGHUP, uncaught exceptions) skipping editor draft saves, session shutdown events, and background job cleanup.
- Fixed /quit and /exit commands blocking session closure by introducing a shutdown budget and backgrounding remaining tasks.
- Fixed git and GitHub CLI subprocesses hanging on interactive prompts by forcing non-interactive environments, adding timeouts, and capping output.
- Fixed RPC mode abort_bash being blocked by running bash commands by dispatching bash in the background.
- Fixed task.maxConcurrency and task.maxRecursionDepth limits being bypassed by sub-spawn paths, ensuring limits are dynamically resized and respected.
- Fixed the edit tool inflating session files by pruning extremely large file snapshots from tool-result details.
- Fixed edit-tool Markdown list guidance so hashline parser errors and the model-facing prompt teach `+- item` escaping instead of steering agents toward full-file `write` fallbacks. ([#4179](https://github.com/can1357/oh-my-pi/issues/4179))
- Fixed workstation OS detection rendering "Kernel: unknown" on macOS 15+.
- Fixed /copy code and /copy cmd commands being treated as normal prompts instead of copying the requested blocks.
- Fixed interactive bash status line not updating after directory changes (cd).
- Fixed session title refreshes ignoring user TITLE_SYSTEM.md overrides during replans, and prevented auto-generated titles from incorrectly preserving all-caps text from user messages.
- Fixed the live todo HUD going stale during long tool-use loops by adding mid-run reminders for incomplete items.
- Fixed /shake and mid-stream chat rebuilds erasing active LLM output.
- Fixed RpcClient failing to restart after being stopped or failing on initial startup.
- Fixed /collab web guests being unable to answer ask tool questions by routing host UI requests through writable collab peers.
- Fixed search and AST tools accepting external read URLs by materializing fetched URL text through the read cache before path resolution.
- Fixed Tavily web search to retry without recency filters if no content is returned.
- Fixed extension validation failures for omp install pi-lean-ctx by exposing legacy tool factories.
- Fixed visibility of the focused option in the multi-select ask picker on certain color themes.
- Fixed TUI row overlapping and duplication in the eval tool's live subagent progress tree under heavy concurrency.
- Fixed session resumes after silent exits by recording pre-tool start markers and shutdown diagnostics.
- Fixed status-line redraw crashes when tool-call arguments contain BigInt values.
- Fixed terminal scrollback rows retaining old colors after theme switches.
- Fixed Esc key behavior in the TUI to clear unrecoverable input instead of preserving drafts.
- Fixed marketplace-installed plugins appearing redundantly in both the npm list and the extension status provider.
- Fixed parent and peer IRC message delivery delays.
- Fixed provider/model:auto entries in modelRoles collapsing to inherit and losing their auto state on reload.
- Fixed plan execution prompts to avoid embedding the full plan, referencing the local plan file instead.
- Fixed subagent live progress leaking raw tool output into the parent TUI.
- Fixed eval subagents with custom output schemas receiving stale incremental yield labels.
- Fixed hidden-thinking live status rows rendering as glyph-only lines by adding a persistent label.
- Fixed /compact summary divider placement to keep it in the live scrollable region.
- Fixed the time_spent status-line segment ticking continuously during idle sessions.
- Fixed browser tool schema validation to require the code argument for run calls.
- Improved robustness of MCP authentication error detection and header-based server discovery.
- Fixed reliable detection of 401/403 authorization failures during Smithery commands and HTTP RPCs.
- Improved streaming preview responsiveness for write, edit, and eval tools by decoding streamed string arguments incrementally.
- Added retry-path diagnostics for assistant-tail removal and scheduled continuations after transient provider errors.
- Fixed CJK history rendering issues across repeated compactions.
- Fixed user-invoked skills failing to identify themselves or resolve relative paths across various execution paths.
- Fixed type errors introduced by the merge sweep: restored the ask row-budget priority field, narrowed dereferenced schema property access, and updated stale test API usage.
- Fixed git clone and fetch being killed by the 5-minute local-command timeout; network transfers now use a separate 30-minute deadline, overridable per call.
- Fixed the TUI collab guest (omp join) silently dropping host ask/selector UI requests; they now present through the standard dialog flow and round-trip responses, with cancellation and resync replay handled.
- Fixed transcript rebuilds (theme change, /shake, focus replay) showing stale streamed write/edit/eval content by sharing the partial-JSON decode between the live streaming path and every rebuild path.
- Fixed an explicitly configured compaction.reserveTokens equal to the built-in default being silently replaced by the proportional small-window fallback; the setting now defaults to unset and explicit values are always honored.
- Fixed user-configured LiteLLM discovery providers keeping stale reseller display-name suffixes for up to 24 hours after upgrade by invalidating the warm model cache.
### Fixed
- Fixed session persistence corrupting Anthropic, OpenAI, and Google signed thinking blocks and reasoning payloads, ensuring cryptographic signatures and encrypted content are preserved verbatim to prevent API replay rejections (HTTP 400).
- Fixed several issues with the `apply_patch` and edit tools, including preventing dirty buffers on early aborts, rejecting overwrites of pre-existing files, stopping at the first failing file in multi-file operations, and pruning extremely large file snapshots to prevent session inflation.
- Fixed process termination (SIGTERM, SIGHUP, uncaught exceptions) to ensure editor drafts are saved, sessions shut down cleanly, and background jobs are cleaned up.
- Fixed `/quit` and `/exit` commands blocking session closure by introducing a shutdown budget and backgrounding remaining tasks.
@@ -89,7 +30,16 @@
- Fixed `/copy code` and `/copy cmd` commands being treated as normal prompts instead of copying the requested blocks.
- Fixed legacy tool compatibility issues for `createReadTool`, `createGrepTool`, and extension validation failures for `omp install pi-lean-ctx`.
- Improved robustness of MCP authentication error detection, Smithery command authorization failures, and streaming preview responsiveness for write, edit, and eval tools.
- Fixed llama.cpp router/preset mode reporting 128k context in the status bar for every preset picked from `/model` regardless of the preset's configured `--ctx-size`. The router-level `/v1/models` only carries `meta.n_ctx` after a preset's child instance is loaded, and router `/props` reports a dummy `n_ctx: 0`, so cold-picked presets fell through to the 128k discovery default. Discovery now also reads `--ctx-size` (or `-c`) from each entry's `status.args` rendered CLI vector and, as a fallback, `ctx-size = N` from `status.preset` INI; the same fallback applies on the runtime refresh that runs when `/model` switches to a preset ([#4190](https://github.com/can1357/oh-my-pi/issues/4190)).
- Fixed llama.cpp router/preset mode reporting 128k context in the status bar for every preset picked from `/model` regardless of the preset's configured `--ctx-size`.
- Fixed post-rewind context to tell the agent the checkpoint completed and to make repeat `rewind` calls recover with guidance instead of a bare no-checkpoint error.
- Fixed grep/ast_grep search scopes rejecting `www.` and collapsed-scheme (`https:/host`) URL spellings.
- Fixed ctrl+p role-model cycling getting stuck on one transition and skipping every other role.
- Fixed interactive bash status line not updating after directory changes (cd).
- Fixed session title refreshes ignoring user `TITLE_SYSTEM.md` overrides during replans, and prevented auto-generated titles from incorrectly preserving all-caps text from user messages.
- Fixed the live todo HUD going stale during long tool-use loops by adding mid-run reminders for incomplete items.
- Fixed `/shake` and mid-stream chat rebuilds erasing active LLM output.
- Fixed Tavily web search to retry without recency filters if no content is returned.
- Fixed user-configured LiteLLM discovery providers keeping stale reseller display-name suffixes for up to 24 hours after upgrade by invalidating the warm model cache.
## [16.2.13] - 2026-07-01
packages/coding-agent/src/session/session-persistence.ts
+28 -17
View File
@@ -59,9 +59,15 @@ function shouldExternalizeImagePayload(
return (key === TEXT_CONTENT_KEY && isImageBlock(value)) || key === "images";
}
/** True for a non-empty string — marks signature/encrypted fields whose block must persist verbatim. */
function isNonEmptyString(value: unknown): value is string {
return typeof value === "string" && value.length > 0;
}
/**
* Recursively truncate large strings in an object for session persistence.
* - Truncates any oversized string fields (key-agnostic)
* - Truncates oversized string fields (key-agnostic), except signed/encrypted
* blocks and signature keys, which persist verbatim
* - Externalizes oversized image payloads to blob refs
* - Updates lineCount when content is truncated
* - Returns original object if no changes needed (structural sharing)
@@ -76,20 +82,23 @@ function truncateForPersistence(obj: unknown, blobStore: BlobStore, key?: string
if (shouldExternalizeImagePayload(obj, key)) {
return { ...obj, data: externalizeImageDataSync(blobStore, obj.data, obj.mimeType) };
}
// Signed/encrypted reasoning is bound to its exact bytes: a truncated `thinking`
// no longer matches its signature and a truncated `redacted_thinking` blob is
// undecryptable, so the provider 400s the replay. Persist these verbatim — never
// truncate, externalize, or descend. Unsigned thinking (e.g. an interrupted
// stream) has no such binding and stays truncatable for size control.
// Signed content is bound to its exact bytes: a truncated `thinking`/`text`/
// `arguments` no longer matches its signature and a truncated
// `redacted_thinking` blob is undecryptable, so the provider 400s the replay.
// Persist signed blocks verbatim — never truncate, externalize, or descend.
// Unsigned blocks (e.g. an interrupted stream) have no such binding and stay
// truncatable for size control.
if (typeof obj === "object" && "type" in obj) {
const signedThinking =
obj.type === "thinking" &&
"thinkingSignature" in obj &&
typeof obj.thinkingSignature === "string" &&
obj.thinkingSignature.length > 0;
const redacted =
obj.type === "redactedThinking" && "data" in obj && typeof obj.data === "string" && obj.data.length > 0;
if (signedThinking || redacted) return obj;
const signed =
(obj.type === "thinking" && "thinkingSignature" in obj && isNonEmptyString(obj.thinkingSignature)) ||
(obj.type === "text" && "textSignature" in obj && isNonEmptyString(obj.textSignature)) ||
(obj.type === "toolCall" && "thoughtSignature" in obj && isNonEmptyString(obj.thoughtSignature));
const redacted = obj.type === "redactedThinking" && "data" in obj && isNonEmptyString(obj.data);
// OpenAI Responses reasoning items (providerPayload.items) carry
// `encrypted_content`, server-validated on replay — atomic like signed blocks.
const encryptedReasoning =
obj.type === "reasoning" && "encrypted_content" in obj && isNonEmptyString(obj.encrypted_content);
if (signed || redacted || encryptedReasoning) return obj;
}
if (typeof obj === "string") {
@@ -97,10 +106,12 @@ function truncateForPersistence(obj: unknown, blobStore: BlobStore, key?: string
return externalizeImageDataUrlSync(blobStore, obj);
}
if (obj.length > MAX_PERSIST_CHARS) {
// Cryptographic signatures must be preserved exactly or cleared entirely — never truncated.
// Truncation would produce an invalid signature that the API rejects.
// Defensive: signature keys normally sit on blocks the guard above returns
// verbatim, but if one is reached here (unknown carrier shape), preserve it —
// truncation produces an invalid signature the API rejects, and clearing
// drops reasoning context the provider needs on replay.
if (key === "thinkingSignature" || key === "thoughtSignature" || key === "textSignature") {
return "";
return obj;
}
const limit = Math.max(0, MAX_PERSIST_CHARS - TRUNCATION_NOTICE.length);
return `${truncateString(obj, limit)}${TRUNCATION_NOTICE}`;
packages/coding-agent/test/session-manager/signature-persistence.test.ts
+2 -37
View File
@@ -27,42 +27,6 @@ function getAssistantMessage(session: SessionManager): AssistantMessage {
}
describe("SessionManager signature persistence", () => {
it("clears oversized signatures instead of truncating them", async () => {
using tempDir = TempDir.createSync("@pi-session-signature-persistence-");
const session = SessionManager.create(tempDir.path(), tempDir.path());
session.appendMessage({ role: "user", content: "continue", timestamp: 1 });
session.appendMessage({
role: "assistant",
content: [
{ type: "thinking", thinking: "reasoning", thinkingSignature: "s".repeat(600_000) },
{ type: "text", text: "done", textSignature: "m".repeat(600_000) },
{ type: "toolCall", id: "tool_1", name: "read", arguments: {}, thoughtSignature: "t".repeat(600_000) },
],
api: "openai-responses",
provider: "openai",
model: "gpt-5-mini",
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,
} satisfies AssistantMessage);
await session.flush();
const reloaded = await SessionManager.open(session.getSessionFile()!);
const assistant = getAssistantMessage(reloaded);
expect(assistant.content[0]).toMatchObject({ type: "thinking", thinking: "reasoning", thinkingSignature: "" });
expect(assistant.content[1]).toMatchObject({ type: "text", text: "done", textSignature: "" });
expect(assistant.content[2]).toMatchObject({ type: "toolCall", id: "tool_1", thoughtSignature: "" });
});
it("externalizes provider image data URLs and restores preserved history payloads across reload", async () => {
using tempDir = TempDir.createSync("@pi-session-provider-image-persistence-");
const session = SessionManager.create(tempDir.path(), tempDir.path());
@@ -265,7 +229,8 @@ describe("SessionManager signature persistence", () => {
it("drops a reasoning signature duplicated by the provider payload and keeps the payload on reload", async () => {
using tempDir = TempDir.createSync("@pi-session-reasoning-dedup-e2e-");
const session = SessionManager.create(tempDir.path(), tempDir.path());
const encrypted = "ENCRYPTED_REASONING_BLOB_UNIQUE_TOKEN";
// >MAX_PERSIST_CHARS: regresses persistence truncating providerPayload reasoning items.
const encrypted = `ENCRYPTED_REASONING_BLOB_UNIQUE_TOKEN_${"E".repeat(600_000)}`;
const reasoning = { type: "reasoning", id: "rs_1", encrypted_content: encrypted };
session.appendMessage({ role: "user", content: "continue", timestamp: 1 });
packages/collab-web/CHANGELOG.md
+2 -2
View File
@@ -5,8 +5,8 @@
### Fixed
- Fixed missing response controls for "ask" questions in the mobile collaboration web UI.
- Fixed an issue where re-sending the same editor "ask" request would clear a guest's in-progress draft response.
- Fixed infinite retry loops in the agent transcript drawer when encountering terminal errors, ensuring the error is displayed and polling stops.
- Fixed an issue where re-sending an editor "ask" request would clear a guest's in-progress draft response.
- Fixed infinite retry loops in the agent transcript drawer by ensuring terminal errors are displayed and polling stops.
- Fixed a delay in displaying pre-welcome connection errors (such as protocol version rejections), allowing the session to terminate immediately with the host's error reason.
## [16.2.0] - 2026-06-27
packages/hashline/CHANGELOG.md
+1 -1
View File
@@ -4,7 +4,7 @@
### Changed
- Optimized stale-anchor remap validation from quadratic to linear complexity, significantly improving performance on large files.
- Significantly improved performance on large files by optimizing stale-anchor remap validation.
### Fixed
packages/natives/CHANGELOG.md
+1 -1
View File
@@ -4,7 +4,7 @@
### Added
- Added workingDir to ShellRunResult to allow hosts to synchronize the session's current working directory without executing a hidden probe command.
- Added `workingDir` to `ShellRunResult` to allow hosts to synchronize the session's current working directory without executing a hidden probe command.
### Fixed
packages/tui/CHANGELOG.md
+1 -1
View File
@@ -4,7 +4,7 @@
### Fixed
- Fixed a potential event loop hang caused by processing oversized, unterminated terminal escape sequences (OSC/DCS/APC).
- Fixed a potential event loop hang when processing oversized, unterminated terminal escape sequences (OSC/DCS/APC).
- Fixed an issue where large Windows terminal session restores could get truncated mid-frame during ConPTY full-paint resume.
## [16.2.13] - 2026-07-01
packages/wire/CHANGELOG.md
+2 -2
View File
@@ -4,11 +4,11 @@
### Breaking Changes
- Upgraded the collaboration protocol (COLLAB_PROTO) to version 3. Guests using version 2 are now rejected during the handshake with a protocol-mismatch error due to new interactive UI request/response requirements.
- Upgraded the collaboration protocol to version 3. Guests using version 2 will now be rejected during the handshake with a protocol-mismatch error.
### Added
- Added collaboration UI request and response frames, enabling browser guests to respond to interactive prompts initiated by the host.
- Added support for interactive UI request and response frames, enabling browser guests to respond to prompts initiated by the host.
## [16.1.8] - 2026-06-20