Responses-style providers serialize providerPayload history items instead of the
visible message blocks when replaying native history. Include providerPayload in
the append-only per-message digest so payload-only history rewrites stop the
stable-prefix walk and re-sync the changed message before any later divergent
tail.
Add a regression where an assistant message keeps identical visible content and
id but changes its openaiResponsesHistory providerPayload while a later message
also diverges; syncMessages must preserve the prefix before the assistant and
refresh the assistant payload.
Fixes#3406
Direct callers can clear AppendOnlyContextManager.log without resetting the
private sync cursor. The advisor reset path does this when recycling its helper
agent, leaving lastSyncCount and messageDigests describing the old transcript
while the physical log is empty.
Clamp the stable-prefix reuse count to the current log length before truncating
and appending. A direct log clear now forces the next sync to replay from index 0
instead of starting from a stale private cursor and dropping prefix messages from
the provider context.
Add a regression that clears the public log after syncing two messages, then
resyncs a context with the same first message and a rewritten second; both
messages must be present in the rebuilt append-only log.
Fixes#3406
Track internal tool-result metadata in append-only per-message digests so
metadata-only rewrites of toolCallId, toolName, or isError stop the stable-prefix
walk and re-sync the changed tool result before any later divergent tail.
This prevents stale tool-result pairing or error state from being preserved when
the text content stays unchanged but provider-serialized metadata changes.
Fixes#3406
`AppendOnlyContextManager.syncMessages` hashed a single rolling digest
over the entire synced prefix, so any in-place rewrite of an already-
synced message — per-turn `pruneSupersededToolResults` / `pruneToolOutputs`
collapsing a tool result, image stripping, or a `transformContext` re-render
— triggered `log.clear()` and re-appended the full conversation from
the current (mutated) view. The provider's cached bytes still matched
the prefix, but every position past the divergence had to be re-prefilled.
On llama.cpp / Ollama / LM Studio this re-prefilled tens of thousands
of tokens every few turns (`n_past \u2248 end-of-system-prompt` collapse,
~40k-token full re-prefill, GPU pinned >400W).
Replace the rolling digest with per-message digests in `#messageDigests`,
walk the new sync against them to find the longest byte-stable prefix,
truncate the log down to that prefix via a new `AppendOnlyLog.truncate(count)`,
and only re-append the diverged tail. Genuine compaction (`length <
lastSyncCount`) still clears the log.
- Tail-only rewrite: prefix stays byte-stable; only the trailing message
re-syncs.
- Deep rewrite: prefix up to the divergence stays byte-stable; the
provider re-prefills from the divergent message onward (architectural
minimum).
- True compaction: unchanged, full replay.
Replace the now-misleading `detects in-place rewrite of already-synced
messages` / `detects in-place rewrite via digest mismatch` tests with
`preserves the byte-stable prefix when a deep message is rewritten (#3406)`,
`preserves the prefix when the tail is rewritten (#3406)`, `appended
new messages keep the prefix stable even when the prior tail also
diverged (#3406)`, and `rewriting the first message still re-syncs
from scratch` so each invariant is asserted directly.
Fixes#3406
- Replaced global timer mocks with local `YieldGate` instances to avoid test flakiness from concurrent environment interference.
- Introduced an injected clock and counting sleep pattern to test gating logic without relying on `process` globals.
- Added a test case to ensure the gate correctly handles negative clock jumps without stalling.
- Extracted yield logic into a configurable YieldGate class to avoid process-global state.
- Injected time and sleep dependencies to support deterministic testing.
- Handled potential negative time progression by forcing a re-anchor instead of gating indefinitely.
- Maintained existing behavior for the public yieldIfDue export via a shared instance.
- Prevented infinite blocking in `yieldIfDue` caused by backward system clock adjustments.
- Implemented time-delta re-anchoring to detect and recover from negative clock drift.
- Decoupled the yield gate into an injectable component to allow deterministic testing.
- Introduced `generateHandoffFromContext` to enable provider-aware oneshot generation and improved cache hit rates via the live-turn pipeline.
- Updated `buildSideRequestContext` to support pinning custom system prompts, preventing per-turn hook leakage during handoff.
- Added concurrency guards across CLI and RPC modes to block manual `/handoff` requests while a session is actively streaming.
- Standardized handoff execution to force `toolChoice: "none"` and enforce consistent cache-routing behavior.
- Replaced `partial-json` with a high-performance `RelaxedJson` parser capable of handling LLM-specific malformations.
- Enabled native support for unquoted keys, single quotes, comments, and Python literals (`True`/`False`/`None`).
- Improved streaming robustness by rejecting non-finite numbers (NaN/Infinity) and rolling back incomplete trailing tokens.
- Optimized tool-call execution by ensuring malformed or truncated arguments are skipped instead of surfaced as undefined/NaN.
- Added comprehensive test suite for both strict repair parsing and partial streaming scenarios.
- Implemented `normalizeAnthropicTargetToolCallId` to define consistent ID validation and fallback logic.
- Integrated the normalization utility into the `transformMessages` function to ensure API compatibility.
- Refactored `transformMessages` to decouple mapping logic from message loop execution for better maintainability.
- Updated the changelog to reflect the correction of tool call ID handling for Anthropic-compatible models.
- Updated the test configuration to specify an explicit compaction strategy.
- Enabled context-full compaction to match expected behavior during tests.
- Updated `render`, `renderMany`, and native snapcompact methods to return promises, ensuring scalable async execution.
- Refactored `transformProviderContext` and `buildSideRequestContext` to support asynchronous operations in agent loops.
- Integrated `Promise.all` for improved concurrency when processing frame rendering and rendering batch operations.
- Updated all internal call sites, SDK hooks, and test suites to accommodate the asynchronous API signatures.
- Verified that side-channel turns maintain stable prompt cache parity with main turns.
- Applied secret and tool obfuscation to ephemeral assistant side-channel data to prevent leakage.
- Updated `Agent` and `AgentSession` test suites to validate consistency under native dialect environments.
CI caught that flooring by the raw local estimate falsely triggers compaction on
thinking-heavy turns: estimateTokens counts the opaque thinkingSignature /
redactedThinking payloads (providers bill them on replay, #2275), but their local
byte size diverges wildly from what the provider actually charges — so a turn
with a large encrypted-reasoning blob but small provider usage would trip the
floor (broke agent-session-handoff 'provider-anchored usage' test).
estimateTokens now takes { excludeEncryptedReasoning } and the compaction floor
(#estimateStoredContextTokens) uses it: the floor counts only reliably-countable,
on-wire-compressible content (text, tool results, tool calls), while the provider
usage arm of compactionContextTokens still accounts for encrypted reasoning. This
keeps the encrypted-reasoning case provider-anchored while still flooring upward
when a before_provider_request hook compresses tool results.
A before_provider_request extension (a context-compression proxy like Headroom,
an obfuscator, or inline snapcompact) can shrink the outgoing request below the
real stored conversation. The provider then reports deflated prompt tokens, so
the auto-compaction threshold never fires and the stored history grows unbounded
until it overflows the context window and can no longer be compacted at all.
Add compactionContextTokens(provider, storedEstimate) = max(provider, estimate)
and apply it to both the pre-prompt and post-response compaction decisions,
flooring the provider-reported tokens by the agent's own estimate of the stored
conversation. Display and cost accounting still use exact provider usage; only
the compaction trigger takes the floor.
- Added `buildSideRequestContext` to the `Agent` class to generate prompt-cache-friendly provider contexts.
- Updated ephemeral side-channel turns to forward the full tool catalog to maintain prompt cache hit rates.
- Injected a `developer` role reminder into ephemeral turns to instruct the model to suppress tool calls.
- Implemented automatic post-processing to strip any tool calls from ephemeral turn responses.
- Exported message and dialect helper functions in `agent-loop.ts` to support context construction.
- Added support for "Fast" serving-path variants for select Fireworks models.
- Updated compatibility logic to route `-fast` suffixes to the appropriate router wire format.
- Extended the model generation catalog to include these Fast variants with their respective pricing.
- Updated AI types to allow the `priority` service tier for Fireworks providers.
normalizeTools only ran toolWireSchema/zodToWireSchema/arkToWireSchema
inside the if (doInjectIntent) branch, so any tool whose intentMode
resolved to "omit" (function intent, or intent: "omit" like the
builtin eval/resolve) had its live arktype/zod schema object left in
parameters. The pruneDescriptions branch already wire-encoded
unconditionally, so the two branches disagreed; first-party providers
re-encoded with toolWireSchema themselves and never noticed, but any
downstream consumer that serialized parameters as-is ended up with an
invalid schema (arktype AST: object-valued required, domain instead of
type, no properties).
Hoist toolWireSchema(t) out of the inject-intent branch so it always
runs and inject intent on top, mirroring the pruneDescriptions shape.
Drop the now-redundant isZodSchema/isArkSchema/zodToWireSchema/
arkToWireSchema imports.
Fixes#3074
- Replaced usage of `ReturnType<typeof setTimeout>` and `ReturnType<typeof setInterval>` with the explicit `Timer` type across the codebase.
- Updated several type definitions and function signatures to use concrete types instead of inferred return types for improved clarity and maintainability.
- Implement JSON repair and strict argument validation to sanitize raw payloads and redact sensitive information from agent event logs.
- Add automatic authentication fallback for benchmark model resolution to ensure consistent performance testing across providers.
- Refactor search tool API parameters by replacing `i` with a case-sensitive `case` boolean flag for clarity.
- Update session history formatting to ensure empty objects are consistently serialized as `{}` instead of empty strings.
- Renamed the global `INTENT_FIELD` constant from `_i` to `i`.
- Updated documentation strings, type annotations, and test expectations across packages to reflect the new field name.
- Ensured consistent usage of the constant in tool schema construction and intent serialization.
- Fixed `SYSTEM.md` integration to correctly include custom-rendered sections like rules and skills.
- Consolidated system prompt validation by requiring `<skills>` tag presence instead of specific prose.
- Removed redundant system prompt math-formatting tests and orphaned task batch documentation tests.
- Prevent object reference sharing between agent snapshots and stream events by deep-cloning tool-call arguments.
- Stabilize GFM tables and Mermaid diagrams during streaming by delaying transcript block commits until content finalization.
- Implement session resume safety to prevent crashes when working directories are missing.
- Add comprehensive test suites to verify streaming commit stability and immutable snapshot isolation.
- Optimized `snapshotAssistantMessageEvent` to accept an optional pre-computed partial snapshot.
- Eliminated redundant deep-clones of `partialMessage` during streaming updates by sharing a single immutable snapshot between the message and the event.
- Removed internal `cloneUnknown` and `cloneToolArguments` functions.
- Switched to `structuredCloneJSON` from shared utils for snapshotting tool call arguments.
can1357