Addresses PR #3412 review (roboomp blocking + codex P2 + Copilot nits):
- The overflow/incomplete retry no longer reuses the COMPACTION_RECOVERY_BAND
hysteresis. Reusing it required residual context below 0.8×threshold, which
turned a recoverable overflow (e.g. 150k on a 200k window, under the ~170k
threshold) into a manual dead-end. Add #compactionCreatedRetryFit, which
measures the rebuilt prompt against the usable fit budget
(contextWindow - effectiveReserveTokens) and is evaluated AFTER the failed
assistant is dropped, so the just-failed turn is excluded.
- The threshold auto-continue keeps the stricter recovery-band check
(#compactionCreatedHeadroom) — that path is the snapcompact thrash guard.
- Split the no-progress warning so each path warns on its own signal.
- Alias errorIsFromBeforeCompaction as assistantPredatesCompaction in the
threshold-usage path for clarity; de-duplicate the rationale comment.
Adds two regression tests (recoverable overflow retries; non-fitting overflow
pauses + warns), mutation-verified against the band-vs-fit split.
(cherry picked from commit f8cf45b2740f04d8776a84521ba563b779d65b25)
prepareCompaction keeps the most-recent turn verbatim (findCutPoint never
cuts at tool results), so when that single turn already exceeds the
compaction threshold the rewritten context stays above threshold. The
context-full / snapcompact auto-compaction success tail scheduled the
agent-authored auto-continue (and the overflow/incomplete retry)
unconditionally, so the next agent_end re-entered #checkCompaction over
the same oversized tail and re-fired forever.
This is the residual loop left after #3247 capped snapcompact's own frame
projection: once #computeSnapcompactMaxFrames drops the frame cap below
one frame, snapcompact is skipped and the context-full summarizer path
still creates no headroom, so the loop persists on that path.
- #runAutoCompaction now gates the continuation and the overflow/
incomplete retry on a post-maintenance headroom check
(#compactionCreatedHeadroom), reusing the shake recovery-band
hysteresis from #2275 (consolidated as the shared
COMPACTION_RECOVERY_BAND). When a pass frees too little it pauses
automatic maintenance and emits a single warning instead of looping.
- The post-turn threshold check ignores an assistant's stale
pre-compaction usage so the scheduled auto-continue can't re-trip on
the kept assistant's old high token count.
Adds a regression test covering the no-headroom (pause + warn, no
continuation) and headroom (auto-continue, no warn) paths; both
assertions are mutation-verified against the guard.
(cherry picked from commit 6fcbbe2b3075827ee621acaee8efa2c586eedd3e)
Stopped snapcompact preflight failures from falling through to the provider-backed LLM summarizer and covered manual plus auto compaction paths.
Fixes#3599
Passed the settled assistant message into session_stop emission so refusal-as-error turns can be pruned from replay context without hiding their stop details from extension hooks.
Expanded the refusal regression test to assert the session_stop payload still exposes the refusal as last_assistant_message.
Fixes#3591
Removed the early return after refusal pruning so the agent_end tail still reaches `#emitSessionStopEvent`, restoring `session_stop` extension hooks (block/continue/telemetry) for refusal-as-error stops.
Regression test wires an extensionRunner with a session_stop handler and asserts it fires for both the refusal turn and the following clean turn.
Fixes#3591
- Added tree-sitter markdown support to resolve headings into full sections in `pi-ast`.
- Enabled block operations (`SWAP.BLK`, `DEL.BLK`, `INS.BLK.POST`) on markdown headings so they encompass the entire section, including nested deeper headings.
- Updated system prompt to guide agents in using structured markdown heading edits for plans.
- Fixed `plan-mode-guard` to correctly resolve local protocol options for subagents.
- Fixed stale `preserveData.snapcompact` frames leaking into context-full compaction after switching from `snapcompact` to `context-full` strategy, which inflated context usage and made sessions appear to compact prematurely.
- Added secret redaction for migrated snapcompact archive plaintext (`text`/`textHead`/`textTail`) during the snapcompact->context-full transition, while preserving opaque provider-replay state byte-identical.
- Added `archiveSourceText()` and `stripPreservedArchive()` utilities to snapcompact module for archive extraction and cleanup.
- Consolidated duplicate `stripSnapcompactPreserveData` functions into `snapcompact.stripPreservedArchive`.
- Added unit tests to verify archive removal and empty state collapse behavior.
The advisor system prompt told the watcher model "at most one advise per
update" and "NEVER send the same advice twice", but nothing enforced
either rule. Issue #3520 captured a session where the advisor emitted
309 advise() calls covering 92 unique notes - 114x "Stop.", 52x "No
issue; continue.", 41x "Done." - landing 309 <advisory severity="blocker">
injections in the primary transcript and destabilizing the watched agent
after the task was already complete.
New AdvisorEmissionGuard sits on AgentSession#enqueueAdvice and:
- Normalizes notes (lowercase, NFKC, punctuation->space, trim) so every
"Stop.", "*Stop*", "STOP!" variant keys to the same canonical form.
- Drops a small allowlist of content-free self-talk filler (stop, done,
complete, no issue continue, lgtm, nothing to add, no further input,
carry on, ...) - silence is the correct expression of "no concerns".
- Dedupes by exact normalized text across the session, FIFO-bounded at
4096 entries.
- Rate-limits to one accepted advise per advisor model prompt cycle. The
runtime calls host.beginAdvisorUpdate?.() before each agent.prompt(),
so the new batch starts with a fresh budget. Suppressed calls don't
consume the budget - a noise call never displaces a real concern.
Reset on advisor reset (compaction, session switch, /new) so a re-primed
reviewer can re-raise old concerns against the rewritten transcript.
Suppression is invisible to the advisor model: AdviseTool still returns
"Recorded." for a dropped call. Surfacing "suppressed" risks the model
rephrasing the same useless note ("Stop." -> "Halt." -> "Cease.") to
bypass the dedupe.
Fixes#3520
Cleared delivered-note memory when the advisor session state resets across conversation boundaries.
Added coverage that repeated advice is allowed again after the dedupe state resets.
Fixes#3511
- Introduced `serviceTierSubagent` and `serviceTierAdvisor` settings to allow independent service tier control for subagents and the advisor model.
- Enabled `"inherit"` mode for these settings, allowing subagents and the advisor to track the main session's live effective service tier, including dynamic toggles like `/fast`.
- Added a resolution layer to ensure service tier propagation from parent sessions to spawned task agents and evaluators.
The preserveCompaction abort path skipped abortCompaction() entirely,
so a manual /compact starting while auto-compaction was in flight no
longer cancelled it. Both passes could then appendCompaction/
replaceMessages, double-rewriting history (reachable via the RPC/
extension compact paths, whose only guard checks #compactionAbortController).
Preserve the just-installed manual controller but still abort the
auto-compaction controller. Adds regression coverage.
Adding `paths` to the global PRIMARY_ARG_KEYS hid the pattern for the
structural tools: ast_grep ({pat,paths}) and ast_edit ({ops,paths})
rendered scope-only (e.g. `ast_grep(src/**/*.ts)`), dropping `pat`/`ops`
— the most decision-relevant argument. Drop the global `paths` key and
special-case `find` (mirroring `search`) so find/search still surface
scope while ast_grep/ast_edit keep showing their pattern via the
existing fallback. Adds regression tests for both structural tools.
Install the manual compaction abort controller before abort teardown so input routing observes session.isCompacting during the starting window.
Fixes#3485
Added scoped path summaries for find/search tool calls in concise session history rendering, with regression coverage for JSON fallback and hidden search scope.
Fixes#3482
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
Trailing empty assistant 'stop' arriving after a successful 'yield'
revived the already-yielded subagent. AgentSession.agent_end maintenance
compared #assistantEndedWithSuccessfulYield(msg) against the trailing
empty-stop message — not the yield-bearing one — so the empty-stop
recovery path appended a retry reminder and scheduled agent.continue().
Track a sticky #yieldTerminationPending flag set when the yield tool
finishes without error and cleared on the next #promptWithMessage. The
agent_end routing extends the existing successful-yield branch: when the
flag is set, or the current message ended with yield, short-circuit
empty-stop / unexpected-stop / compaction continuations for the rest of
the run, so a successful yield is terminal regardless of trailing stops.
Fixes#3389
GitHub Copilot's /models response advertises supports.vision = true for
Claude/GPT chat models on every host, but only the canonical personal
endpoint (https://api.githubcopilot.com) actually accepts image inputs;
the business (api.business.githubcopilot.com) and enterprise
(copilot-api.{domain}) hosts respond '400 vision is not supported'.
snapcompact then injected rasterized transcript frames after compaction
and permanently broke every business-Copilot session.
- Catalog discovery (githubCopilotModelManagerOptions.mapModel) now
forces input=['text'] whenever the resolved baseUrl is not the
canonical personal-Copilot host, so the upstream's vision flag is
honoured only where it actually works.
- mergeDynamicModel honours the dynamic input value (instead of
OR-upgrading with the bundled reference) when the merged baseUrl
differs from the bundled one, so a bundled spec pinned to the
personal host can no longer taint a business-resolved merge.
- snapcompact-inline's canSendImages helper short-circuits the
rasterizer for any github-copilot model whose baseUrl is non-personal,
catching stale cached specs that still advertise vision.
- Helper isPersonalGitHubCopilotBaseUrl exported from
pi-catalog/wire/github-copilot so catalog and coding-agent share one
canonical check.
Regression coverage in github-copilot-model-limits.test.ts (vision
endpoint policy + full merge) and snapcompact-inline.test.ts (#3387
business/enterprise case).
Fixes#3387
With collapseCompactedHistory the live display fell into the LLM compaction
branch, which skips the firstKeptEntryId..compaction turns whenever an OpenAI
remote-compaction replacementHistory payload is present. That payload feeds the
provider only and is not rendered, so a remotely-compacted session showed just
the summary plus post-compaction rows, hiding recent turns that were visible
before. Emit the kept SessionEntry rows in transcript mode regardless. Adds a
regression.
Devin provider models (devin-agent) advertise reasoning: true but no
thinking.efforts metadata — Cascade selects effort by routing to sibling
model ids, not a wire param. getSupportedEfforts(model) therefore returns
[]. clampAutoThinkingEffort previously short-circuited that empty supported
list by returning the requested effort as-is, so the auto-thinking
classifier-resolved level (e.g. low) reached stream.ts:1163 where
requireSupportedEffort threw 'Thinking effort low is not supported by
devin/<id>. Supported efforts: '. In --print mode the user saw the error
text; in the TUI it was silently swallowed, producing the reported
'working then empty response' symptom.
Returns undefined when supported is empty so the result mirrors
clampThinkingLevelForModel's behavior on the same shape (the explicit
--thinking low / high paths already worked because of this). Updates
classifyDifficulty's return type to Effort | undefined and threads through
to the existing #applyAutoThinkingLevel undefined-effort early-return.
#applyAutoThinkingLevel also short-circuits the classifier call up front
for these models — there is no effort to pick.
Fixes#3356
Read per-model llama.cpp meta.n_ctx values during discovery, refresh selected models after lazy load, and bypass fresh cache reuse for llama.cpp refreshes so server restarts update context windows.\n\nFixes #3310
