- Added a 60-character minimum length threshold for demoting interrupted thinking into hidden continuity context in agent-session.ts.
- Updated LLM message conversion in messages.ts to strip incomplete thinking from user-interrupted assistant turns regardless of continuity note presence.
- Added test coverage in agent-session-interrupted-thinking.test.ts verifying behavior for reasoning lengths below and at the threshold.
Why:
The module-global array memo strongly retains the most recently converted
transcript and output after its session is disposed.
Changes:
- Store exact-repeat and append-growth state in a WeakMap per input array.
- Preserve generation invalidation and per-message weak caching.
Evidence:
- Seven disposal runs collected both arrays after 50 forced-GC passes and
reduced median heap delta by 87.54%.
Refs #8119
Codex found the shared attribution predicate recognising only tool calls,
text and signed thinking. A native image response often arrives with no
text and no tool call at all, so an image-only turn was read as producing
nothing: attribution stayed on whichever model spoke before it, and the
empty-stop rule could classify a successful generation as empty.
Everything the assistant can emit now counts except two: unsigned thinking,
which is not provider-authenticated and was already excluded, and
Anthropic's `fallback` marker, which records that a request was routed
elsewhere rather than carrying output. Redacted thinking and server-tool
blocks are real work by the same argument as the image.
The `toolUse` arm keeps its stricter rule — an orphaned toolUse stop needs
a tool_use block to anchor a later tool_result, and an image cannot.
Also switches the new test to the namespace import AGENTS.md requires for
node builtins.
An Agent Hub row reported a subagent as having run on a model that never
spoke. All 97 of its requests, 421K tokens and $6.19 of cost were served
by the primary; a transient stall then armed a fallback, that fallback
errored on its first request with an exhausted quota, and the run died.
Attribution followed the routing switch rather than the output.
Three surfaces lied independently, each re-deriving "the current model"
and calling it the run's model: the executor's progress snapshot, the
session's fallback selector that the hub row reads first, and the
transcript walk behind a settled row.
Sessions now own attribution. `AgentSession.servingModel` names the model
that produced this session's output, holding the last model that actually
served while a candidate is armed but unproven. A switch is a routing
decision, not evidence the target can produce anything, so the answer only
moves once a turn on the target settles.
Consumers read it instead of reconstructing it. The executor's observer
dropped its own event bookkeeping: that bus also carries advisor turns
running on a different model, and it was reading `retry_fallback_applied`
as proof of service. The hub row reads the same getter, so the main
session — which has no executor progress and no persisted history — stops
rendering an unproven candidate as its plain configured model. A fallback
armed before anything has served is still shown, marked as a fallback,
because there is no earlier work to miscredit there.
One predicate decides "this turn produced output", shared by the live
session and the offline replay so they cannot disagree. `error` and
`aborted` are both failures — a stalled stream is finalized as `aborted`
with its partial block still attached, so a stop reason alone proves
nothing — and a turn needs actionable content, which a `length` stop
burning its budget on unsigned thinking does not have. It tolerates
malformed content blocks: transcripts outlive the shapes that wrote them,
and one bad line previously blanked a whole row's history.
Ordering matters at two swap sites. Both the chain advance and the
cooldown-expiry restore move the model and fan `model_changed` out to
subscribers synchronously, so each now updates fallback state before the
swap rather than after; otherwise an observer reading attribution inside
that window sees the incoming candidate carrying the outgoing one's proof.
A startup-selected fallback owns the run from its first request only on a
fresh session. A resumed transcript already holds turns another model
produced, so there the candidate stays unproven until it answers.
`retryFallbackModel` is removed: every consumer reads `servingModel`, and
keeping a parallel derived getter alive for tests is the duplicate surface
this change set exists to remove.
Routed direct custom-message conversion through the collab steering transform so side requests and compaction see the same enveloped user turn as primary requests.
Extended the regression test to exercise convertToLlm without transformContext.
Converted user-attributed collab prompt frames to prioritized user messages only on the model-facing path, preserving guest details in persisted transcript frames.
Added regression coverage for the provider role, steering envelope, and retained guest attribution.
Fixes#7288
- Add the `app.live.toggle` keybinding defaulted to `Ctrl+L` to start or stop live voice mode.
- Remap the default display-reset action (`app.display.reset`) from `Ctrl+L` to `Alt+L`.
- Update the live visualizer to listen for stop keys so the toggle chord terminates active sessions.
Retained complete native web-search call/result pairs through leaked-thinking projection at their original source anchors.
Kept opaque server-tool blocks atomic during persistence, stripped them on reparent, and covered custom continuations plus compaction retention.
Fixes#6703
- Added `LIVE_DELEGATION_MESSAGE_TYPE` constant and delegation message handling for voice sessions.
- Implemented turn-based transcript coalescing with user and assistant turn counters.
- Added transcript display row with normalized rendering in the live visualizer.
- Refactored controller to send delegation messages via `sendCustomMessage` with configurable frame styling.
- Removed microphone permission error reporting from silence detection logic.
- Extracted internal handlers and logic from AgentSession into dedicated runner, guard, and coordinator modules.
- Created standalone modules for bash execution, evaluation runners, IRC bridging, and prewalk coordination.
- Established dedicated session components for tracking stats, todos, streams, and retry fallback chains.
- Preserved existing session behavior while significantly reducing monolithic class size and complexity.
Long sessions re-walked the full live AgentMessage[] every turn: convertToLlm
re-converted the unchanged prefix and estimateTokens re-tokenized settled tool
results and assistants, redoing work only the newest suffix can change.
- Added a per-message estimate cache in agent-core keyed by identity, with a
settle gate (assistants cache only with real usage + terminal non-error
stopReason; streaming partials bypass) and dual option-split WeakMaps for the
default vs compaction-floor estimates.
- Memoized convertToLlm per message identity + assistant interruptedNext flag,
with an exact-repeat outer-array reuse and slice-on-growth for append-only
turns, guarded by a boundary-identity check against interior splice-replaces.
- Invalidated both caches at the mutation seams: prune, shake, strip-images, and
the prewalk plan-nudge scrub, via invalidateMessageCache /
registerMessageCacheInvalidator across the package boundary.
- Added the llm-assembly bench (N=5000, robust MAD-noise gate): steady/append
convert and repeat estimate are all >10x faster with noise under 20%.
Fixes#5934
Switching from a vision model to a text-only model kept replaying
historical image content blocks to the new provider, which rejected them
with invalid_argument. The convertToLlm wrapper in sdk.ts only filtered
images when images.blockImages was set, never by model capability, so
the outbound request carried image blocks the active model could not
accept.
Add replaceLlmImagesWithText() to scrub image blocks out of the
already-converted LLM message view, and call it from the sdk wrapper
when the active model's input lacks "image". History on disk keeps its
images; only the provider request is scrubbed, and the check reads the
active model dynamically so a /model switch takes effect next turn.
Fixes#5400
A request the provider rejects (e.g. 413 oversized payload) yields a
synthesized assistant turn with empty content and stopReason 'error'.
That turn is written to session.jsonl, so on reload it replays as an
empty assistant turn and re-sends the same rejected context. Keep the
rejection UI-only (pinned error) and out of persisted history so a
reloaded session resumes from the last good turn.
Normalized extension custom-message payloads before session state or persistence, including bare string sendMessage shorthands. Skipped legacy bare custom_message entries during context rebuilds and dropped malformed custom/hook messages before LLM conversion. Added regression coverage for the poisoned-session resume crash.\n\nFixes #4345
- Introduced `isProbablyBinary` utility to sniff file headers for NUL bytes or invalid UTF-8 sequences.
- Updated `ReadTool` to use the binary sniffer, preventing mojibake corruption in output when reading non-text files.
- Refined `file-mentions` auto-reads to skip binary files and mark them as `binary` in the message transcript.
- Added comprehensive unit tests for binary detection logic, covering NUL bytes, truncated multibyte characters, and path-based file sniffing.
- Updated the demotion logic to treat complete, signed thinking blocks as stable content that terminates an interrupted stream.
- Protected signed thinking runs from being stripped when processing interrupted agent messages.
- Updated `convertToLlm` to detect and strip trailing thinking runs from user-interrupted assistant messages.
- Retained original thinking content on the persisted assistant message to ensure UI state (render, reload, and rebuild) remains intact.
- Added `followedByInterruptedThinking` helper to verify continuity message presence before stripping for the provider request.
- Updated persistence tests to confirm thinking remains in the session state but is excluded from LLM context headers.
- Introduced V2 streaming remote compaction for OpenAI-compatible models, enabling full conversation history forwarding and reducing data loss from local trimming.
- Added comprehensive support for sessionId, promptCacheKey, and automatic retry mechanisms to improve compaction reliability and accuracy.
- Updated agent, catalog, and configuration schemas to manage V2 streaming settings, model metadata, and model-specific context window constraints.
- Extended freeform tool patch support for Azure OpenAI and Codex models and refined assistant-side history preservation across providers.
- Added `INTERRUPTED_THINKING_MESSAGE_TYPE` and `demoteInterruptedThinking` in `session/messages.ts` for stripping unfinished thinking into durable context.
- Persisted hidden interrupted-thinking context from `session/agent-session.ts` immediately after user-interrupted assistant turns.
- Added the `prompts/system/interrupted-thinking.md` envelope used for replaying interrupted reasoning.
- Added focused interrupted-thinking coverage in `test/agent-session-interrupted-thinking.test.ts` and `test/session/interrupted-thinking-demote.test.ts`.
Split image-bearing custom messages into developer text plus user image content before provider conversion so queued skill images stay out of developer slots.
Mapped user-attributed skill prompt custom messages to user-role model messages while preserving developer framing for auto-applied skills and other custom messages.
Fixes#3698
- Migrated 288 lines of scattered error classification logic from `utils/error-id.ts` into a cohesive `packages/ai/src/error/` module with 13 specialized submodules covering flags, classes, OAuth, providers, rate-limiting, and finalization.
- Replaced 100+ generic `Error` throws across 60+ provider and registry files with semantic `AIError.*` classes (e.g., `AIError.MissingApiKeyError`, `AIError.OAuthError`, `AIError.ProviderResponseError`), improving error diagnostics and retry logic.
- Consolidated error utility imports from `pi-utils` and scattered classification functions into a single `AIError` namespace, reducing coupling and simplifying error handling across all packages.
Reviewer caught that demoting the whole mixed payload to `user` (`@notes.md
@screenshot.png`) regressed the developer-priority treatment text-only
mentions still get for image-free turns. `generateFileMentionMessages` packs
every `@…` into one `fileMention`, so the previous `hasImage` toggle
collapsed the source-file context into the user slot whenever an image was
attached.
`convertToLlm` now returns up to two messages per `fileMention` via
`flatMap`: text-only files keep their existing `developer` envelope, and
image-bearing files emit a separate `user` envelope that carries their
`<file>` wrappers plus the `input_image` block. Pure-text and pure-image
turns still collapse to a single message.
Tests cover the mixed case (split into developer + user), the image-only case
(single user message), and the existing text-only case (single developer
message).
Fixes#3443
Codex GPT models on chatgpt.com /codex/responses rejected `@image` turns with
`Codex error event: [OneOfParam] [input[N].content[M]] [invalid_enum_value]
Invalid value: 'input_image'. Supported values are: 'input_text'.` —
`convertToLlm`'s `fileMention` arm always emitted a `developer`-role
Responses message, but a developer-role content slot only accepts
`input_text`. #3421's prior fix only suppressed the Codex Responses Lite
header on image-bearing turns; the full transport kept rejecting the same body.
`fileMention` now uses `user` role when any attached file carries an image;
text-only mentions keep `developer` so the auto-read context still rides at
instruction priority for the agent.
Fixes#3443
- Changed `wrapSteeringForModel` to wrap every `steering:true` user message regardless of position, so a steer's wire bytes stay identical once buried instead of reverting to raw and busting the prompt cache.
- Neutralized the present-tense framing in `user-interjection.md` so always-wrapping no longer leaves stale "current task" wording on buried interjections.
- Updated `session-messages.test.ts` to assert buried steering messages are wrapped too.
Empty provider-side aborted turns now enter the existing auto-retry path without switching retry model fallback, while aborted turns with partial content still settle normally.\n\nFixes #2685
- Removed the streaming guard that previously rejected /tan while the parent response was still generating.
- Passed "deliverAs: \"nextTurn\"" when sending the background dispatch breadcrumb and kept "triggerTurn: false" so an in-flight turn is not steered.
- Skipped rebuilding chat messages during streaming sessions and updated tests to cover the non-blocking dispatch path.
- Replaced queued-message interrupt flow with session abort calls on empty submit and escape.
- Removed interrupting state and notifyInterrupting teardown paths from abort handling.
- Updated AgentSession queue operations to use shared steering and follow-up queue views.
- Propagated isAborting through session state and collab payloads to suppress late updates.
Adds compaction.strategy: "snapcompact" to the schema and the AgentSession routing: when chosen, both manual /compact (without custom instructions) and auto compaction call snapcompactCompact() to archive history as PNG frames instead of an LLM summary. Falls back to context-full with a visible warning notice when the current model is text-only or when /compact gets custom instructions. CustomTool and shared-event payloads carry the new action through. \n\nAlso wires the per-turn supersede pass: #pruneSupersededReads() runs every turn before threshold gating (cache-aware: only fires when the post-candidate suffix is small or the prompt cache is cold), prunes older read results superseded by a newer read of the same file, rewrites the session, and accounts the saved tokens in the next compaction decision. Gated by compaction.supersedeReads (default on).\n\nsession/messages.ts now delegates the core role conversion to agent-core's convertMessageToLlm so snapcompact image blocks flow through the LLM-context conversion path.
- Added per-path edit versioning in `EditTool` to drop stale late diagnostics after later edits.
- Added deferred diagnostics queueing through `queueDeferredDiagnostics` and late-diagnostic yield batching.
- Updated `UiHelpers` to render late diagnostic file path and summary lines in the chat transcript.
- Added `isUserInterruptAbort` and `shouldRenderAbortReason` helpers so interrupt handling can distinguish Esc-based aborts from other abort reasons.
- Updated assistant transcript rendering to suppress the `Interrupted by user` line while continuing to show generic or other abort labels.
- Updated plan-review and transcript container tests to assert interrupted assistant messages no longer render the redundant interrupt line.
- Updated `convertToLlm` logic to emit `developer` role for custom, hook, and file-mention inputs.
- Simplified OpenAI compact output filtering to retain only `user` and `assistant` messages, removing legacy `system-reminder` pattern checks.
- Adjusted compaction and session tests to match the new developer-role mapping and expected compacted content.
- Added optional `reason` parameters to `Agent.abort` and `AgentSession.abort` APIs.
- Passed abort reasons through interrupt flows into underlying agent cancellation.
- Replaced hard-coded abort text with `resolveAbortLabel` for streaming and replayed messages.
- Fell back to generic `Request was aborted` text when no abort reason was provided.
- Marked steering user messages and wrapped them pre-LLM so the model sees a `` envelope.
- Kept transcripts and persisted history with the user's original raw text.
- Restored `AssistantMessageComponent` stable-prefix completion API.
- Added `/drop-images` slash command handling for runtime and TUI to drop images and report results.
- Added `stripImagesFromMessage` utilities to remove image blocks from message content and return removal counts.
- Added `AgentSession.dropImages()` to prune image blocks, rewrite history when needed, and rebuild session context.
- Added tests for user, toolResult, fileMention, and assistant image-stripping, placeholders, and zero-removal cases.
- Relocated compaction, branch-summarization, pruning, and utils from coding-agent to packages/agent/src/compaction.
- Moved OpenAI remote compaction helpers from packages/ai to the new compaction module.
- Added handoff.ts with extractHandoffDocument, createHandoffContext, and renderHandoffPrompt helpers.
- Exposed new entries.ts with standalone SessionEntry types so coding-agent no longer owns them.
- Removed export leakage by demoting many helper and const symbols to module-local scope.
- Renamed underscore-prefixed internals and cache fields, then updated related references and `satisfies never` checks.
- Deleted obsolete logic branches and helpers, including harmony-stream interruption flow and unused benchmark runtime helpers.
- Updated Biome config and manifests by broadening lint coverage and removing an unused `@napi-rs/cli` dev dependency.
- Adjusted tests and utilities to use renamed test helpers and remove redundant private test-only helpers/locals.
Codex review flagged that the silent-abort sentinel
("__omp.silent_abort__") persists into AssistantMessage.errorMessage
but three downstream consumers render errorMessage verbatim:
- session-observer-overlay.ts: renders "✗ Error: __omp.silent_abort__"
when content is empty (confirmed user-visible today)
- print-mode.ts: writes marker to stderr and exits non-zero (latent;
plan-mode→compact not reachable from print mode today, but unguarded)
- acp-agent.ts: emits marker as agent_message_chunk text to ACP
clients when message has no other notifications (latent)
Add isSilentAbort() guard at each site. Extend the SILENT_ABORT_MARKER
consumer list in messages.ts doc comment to include all six consumers.
Add regression tests: overlay (2 tests), print-mode (2 tests), ACP
replay (1 test).
Op: correct
Restores: spec:silent-abort-marker-never-surfaces
can1357