User-invoked skills (typed /skill:, steered, follow-up, interrupted/
resumed via compaction, ACP, RPC) only appended a bare "Skill: <path>"
line, so the model neither learned the user had invoked that specific
skill nor where the skill directory was. Relative paths in skill bodies
(scripts/, templates/) could not be resolved.
Route all user-invoked paths through a self-identifying, baseDir-aware
prompt template; keep hidden autoload skills on the minimal non-user
format. Interactive skillCommands now carries the loaded Skill object
instead of a bare path so baseDir flows through without reconstruction.
The invocation kind defaults to "user" to keep buildSkillPromptMessage
source-compatible.
Op: correct
Restores: spec:user-invoked-skill-prompt-self-identifies-and-exposes-skill-directory
- Removed prohibitions against using grep, rg, ls, find, head, tail, and redirections in the bash tool.
- Clarified in output specifications that stderr is merged into stdout.
- Added a last-resort recovery step to run `shake("elide")` on oversized message tails when auto-compaction cannot otherwise free enough context.
- Re-tests the context headroom and auto-continue predicates after a successful rescue before falling back to pausing maintenance.
- Updated the dead-end warning message to suggest running `/shake images` for irreducible, image-only tails.
ToolArgsRevealController.setTarget initialized new entries with revealed=0, so the first message_update returned { __partialJson: "" } even when the provider had already parsed a complete chunk. For renderers without exposeRawPartialJson (e.g. write), the throttled re-parse + cached displayArgs short-circuited every subsequent setTarget, leaving the preview body blank until tool_execution_end.
Seed revealed with the full incoming partialJson length on entry creation (clamped to a surrogate-safe boundary). The first frame now carries the parsed path/content immediately; subsequent message_updates extend target and the reveal ticks pace only the newly arrived bytes — no field is ever truncated because the seeded prefix is the longest the entry has seen so far.
Fixes#3881
fix(compaction): cap snapcompact frame payloads (#3866)
Bound rebuilt snapcompact image payloads by a per-request base64 byte
budget so long sessions stop re-sending multi-megabyte standing image
archives on every provider request; auto-compaction falls back to
context-full summaries when snapcompact output is too large.
Resolved snapcompact.ts conflict against the main font-rendering refactor
by keeping both renderabilityProbeText and the frame-budget helpers.
Fixed historyBlocks to emit the omitted-frame notice before the kept
(newer) images, since the byte budget drops the oldest frames — keeping
reconstructed blocks oldest-to-newest (addresses Codex P2 review).
Fixes#3792
- Added eight new Go-specific rules to the discovery package.
- Registered the new Go rules in the default rule source index.
- Covered the new Go AST matching conditions with test cases in `builtin-defaults.test.ts`.
Forwarded persisted provider stream timeout settings into model requests so slow local LLM streams can widen or disable first-event and idle watchdogs without environment variables.
Fixes#3878
Normalized string-encoded JSON arrays in grep/search path handling so direct execute paths match validated tool-call behavior.
Added regression coverage for direct GrepTool.execute paths supplied as a JSON-array-shaped string.
Fixes#3873
The yield tool's per-call schema validator was skipped entirely for incremental
yields (`type: ["<label>"]`), so when a subagent emitted a non-conforming value
for a known section (e.g. DeepSeek-v4-pro returning "Correct"/"correct."/"approved"
for the reviewer's `overall_correctness` enum), the call succeeded locally and
the model got no retry feedback. The mismatch only surfaced post-mortem in
`finalizeSubprocessOutput` as a fatal `schema_violation` — the parent agent
lost the entire result, with no recourse for the subagent to fix it.
Build a per-label sub-validator map alongside the full-schema validator: each
entry validates one section's `data` against its top-level property's sub-schema
(items schema for array-typed labels like `findings`). The yield tool runs this
map for incremental yields and routes failures through the same MAX_SCHEMA_RETRIES
budget the terminal path uses, so the model sees up to three corrective retries
and the existing schema-override safety net accepts the value with
SUBAGENT_WARNING_SCHEMA_OVERRIDDEN after exhaustion. Unknown labels remain
unconstrained so scratchpad/streaming sections still pass.
Fixes#3870
When legacy snapcompact archives exceed the per-request byte budget, retain frames from the newest end of the archived middle and restore oldest-to-newest order for the kept subset.
Apply the snapcompact frame byte-budget cap even when the active model has no known context window, avoiding 80-frame archives on custom vision models.
Bounded persisted snapcompact image archives by base64 byte size so large sessions stop re-sending multi-megabyte frame walls on every provider request.
Auto snapcompact now falls back to context-full summaries when rendered frame payloads exceed the byte budget, and legacy oversized archives omit over-budget frames during LLM context rebuilds.
Fixes#3792
- Updated the default browser User-Agent string to emulate a modern version of Chrome.
- Added typical browser headers to the outgoing fetch request, including Sec-Ch-Ua, Sec-Fetch flags, and Referer.
- Added a blank "b" parameter to the form body to match native DuckDuckGo HTML search behavior.
- Guarded the `all_turns` reasoning context value to OpenAI models version 5.4 or greater.
- Suppressed `reasoning.context` defaults and explicit overrides when `all_turns` is requested on unsupported models to prevent server rejection.
- Introduced `supportsAllTurnsReasoningContext` helper in `@oh-my-pi/pi-catalog/identity` using semver classification.
- Updated request-transformer and response-options logic to conditionalize the request payload shaping.
- Expanded test coverage to verify correct fallback behavior and explicit overrides across gpt-5.x versions.
- Added `leaked-thinking-stream.ts` whose `wrapLeakedThinkingStream`/`LeakedThinkingProjector` re-projects any provider stream, splitting leaked ```thinking`/`<think>` fences out of the visible-text channel into structured thinking blocks live while preserving text/thinking/tool signatures.
- Wrapped the shared `withProviderInFlightLimit` dispatch (both the unlimited and queued paths) and the standalone GitLab Duo path in `stream.ts` so healing covers every provider exit idempotently.
- Reworked `google-gemini-cli.ts` visible-text emission onto `StreamMarkupHealing` with `feedVisibleText`/`flushVisibleText` and explicit block bookkeeping, lifting leaked Gemini fences ahead of native tool calls.
- Added `leaked-thinking-stream` coverage and a gemini-cli healing case, and updated `stream-auth-retry`/`google-gemini-cli-alignment` expectations for the healed event sequence and dropped empty-text residue.
- Recorded the changelog `### Fixed` entry (whose run also carries the adjacent Codex `all_turns` line).
- Added `google-interactions.ts` streaming the Gemini Interactions API (model-mode steps, thought/tool/text deltas, usage, `previous_interaction_id` lineage) shared by the direct Google and Vertex providers.
- Routed `streamGoogle` and `streamGoogleVertex` through `resolveInteractionDispatch`, defaulting Gemini 3+ onto Interactions (official endpoint for direct Google, bearer/ADC for Vertex) with transparent `:streamGenerateContent` fallback and a `useInteractionsApi: false` opt-out.
- Added explicit Vertex bearer support in `google-auth.ts` via `GOOGLE_CLOUD_ACCESS_TOKEN`/`CLOUDSDK_AUTH_ACCESS_TOKEN` plus a `hasVertexBearerCredentialsHint` probe gating the auto-default.
- Added `useInteractionsApi`/`storeInteraction`/`previousInteractionId` to `StreamOptions` and `GoogleSharedStreamOptions`, threading them through `mapOptionsForApi`.
- Added `google-interactions` coverage and pinned the existing generateContent assertions with `useInteractionsApi: false`.
- Reordered `AuthStorage.getApiKey` and its async session variant to resolve stored OAuth, then an env var, then a stored static api_key — so an explicit `GEMINI_API_KEY`-style env var now wins over a stale broker-migrated key.
- Reworked `getCredentialOrigin` and `getApiKeySource` to mirror that precedence via a `describeStored` helper, reporting `oauth`/`env`/`api_key` in the new order.
- Updated `auth-storage-api-key-login` to neutralize ambient env keys with a `getEnvApiKey` spy, and `auth-storage-credential-origin` to assert OAuth outranks api_key and env outranks a stored api_key.
StdinBuffer held a bare `\x1b\x1b` chunk and timer-flushed it as one
sequence. `parseKey("\x1b\x1b")` returns undefined, so CustomEditor
fell through to the base editor and never fired the configured `onEscape` —
the double-escape gesture and the second-press single-Esc handler both went
dead whenever the terminal batched the two presses into one stdin read.
Split an exact bare `\x1b\x1b` into two ESC events only after the
flush window proves no follower arrived. If a follower does arrive, emit the
first ESC and restart parsing at the second ESC so legacy Alt chords
(`\x1bd`, `\x1b\x7f`) remain one downstream keypress. Meta-CSI/SS3
chords (`\x1b\x1b[A`, `\x1b\x1bO…`) still emit as one combined
sequence.
EventController.tool_execution_update re-armed the working loader when a
transient overlay (auto-compaction / auto-retry / handoff) had torn it down
mid-tool; tool_execution_end did not. A subagent (`task`) call only fires
_end, so a task result landing after such an overlay left the UI looking
idle even though the session was still streaming. Mirror the reconciler
call in #handleToolExecutionEnd.
Fixes#3857
`tool_execution_end` for a long-running tool (`task` subagent, async bash
poll, …) is the next streaming event on the parent session when an inner
transient overlay (auto-snapcompact, auto-context-full, auto-retry) nulled the
working loader between the tool's start and end. The overlay-end handlers are
the only loader restorers keyed off the missing reference; if the subagent's
`tool_execution_end` lands while the overlay is still active (or its end
handler errored before re-arming), the spinner stays gone for the rest of the
parent turn even though the agent keeps streaming.
`#handleToolExecutionEnd` now calls `#ensureWorkingLoaderWhileStreaming()`
at the top, mirroring `tool_execution_update` so the working loader survives
a subagent completing inside the overlay window.
Refs #3858
StdinBuffer held a bare `\x1b\x1b` chunk (or emitted it as one when followed by
a non-CSI byte). `parseKey("\x1b\x1b")` returns undefined, so CustomEditor
fell through to the base editor and never fired the configured `onEscape` —
the double-escape gesture and the second-press single-Esc handler both went
dead whenever the terminal batched the two presses into one stdin read.
Split a bare `\x1b\x1b` into two ESC events at the buffer layer, mirroring
the existing split for ESC + SGR mouse report. Meta-CSI/SS3 chords
(`\x1b\x1b[A`, `\x1b\x1bO…`) still emit as one combined sequence.
EventController.tool_execution_update re-armed the working loader when a
transient overlay (auto-compaction / auto-retry / handoff) had torn it down
mid-tool; tool_execution_end did not. A subagent (`task`) call only fires
_end, so a task result landing after such an overlay left the UI looking
idle even though the session was still streaming. Mirror the reconciler
call in #handleToolExecutionEnd.
Fixes#3857
Mapped Cerebras gemma-4-31b dynamic discovery to include image input capability and covered the OpenAI Chat Completions image_url serialization path.
Fixes#3854
- Introduce `fallbackUrl` to allow retrying requests on the global Vertex endpoint when a 404 is encountered on a regional endpoint.
- Correctly track tool call names to facilitate accurate function response mapping.
- Update `serviceTier` type definitions to accurately reflect allowed values.
- Refine vertex endpoint location resolution to ensure better compatibility with ambient regional settings.
Reviewer flagged that the omit/forced-tool gate matched every public id, regressing Fireworks and OpenRouter Kimi K2.7 Code (non-zai dialects).
Added Fireworks + OpenRouter regression coverage.
- Migrated global service tier settings to a per-model-family architecture (OpenAI, Anthropic, Google).
- Implemented `ServiceTierByFamily` mapping to allow independent configuration and resolution per provider.
- Added automatic migration logic for legacy service tier and fast-mode application settings.
- Updated telemetry, session management, and task execution to support provider-specific tier resolution.
Kimi K2.7 Code rejects disabled thinking on native Kimi endpoints, so route caller disable requests through the omit mode and let the model default to required thinking.
Added regression coverage for the title-generator-style Kimi Code request, Moonshot K2.7 Code variants, and K2.6's still-supported disabled-thinking path.
Fixes#3852
metaphorics