`#resolveAdvisorRuntimeDescriptors` hardcoded `ThinkingLevel.Medium` when
no thinking suffix was configured. For reasoning models with no
controllable effort surface (`devin-agent`: `reasoning: true`,
`thinking: undefined` — Cascade selects effort by routing to sibling
model ids, not a wire param), that default tripped
`requireSupportedEffort` on the first advisor prompt with an empty
`Supported efforts:` list, disabling the advisor session-wide.
Route the default through `resolveThinkingLevelForModel(model, level)`
which preserves explicit `off`, clamps a concrete effort into the
model's supported range, and returns `undefined` for reasoning models
without controllable efforts — falling back to `Inherit` so no effort is
sent while reasoning stays enabled. Matches the `auto`-path fix
(`clampAutoThinkingEffort`) and the Autonomous Memory clamp
(`clampThinkingLevelForModel`).
Fixes#4579
- Introduced an automated retry recovery system to track, manage, and persist recovered error states within agent sessions.
- Enabled compact transcript rendering for recovered auto-retry errors by removing heuristic commit machinery.
- Improved raw read tracking and provenance in the ReadTool to support refined file snapshot recording and hashline editing.
- Excluded recovered assistant messages from default model context and updated event controllers to handle retry recovery life cycles.
- Made resolveShapeForText choose silver16-bw for CJK-heavy auto transcripts while preserving explicit variants and unsafe glyph protection.
- Added silver16-bw to the snapcompact shape settings submenu and renamed unsupported-glyph warnings.
- Covered auto shape selection, explicit variant precedence, unsafe glyph scans, and settings option parity.
Fixes#4486
- Added a Usage.orchestration sidecar for provider-side service tokens so Responses/Codex totals and costs stay accurate without inflating visible prompt input/cache buckets.
- Updated Codex/WebSocket usage, session/status aggregates, and usage reporting to preserve orchestration-aware totals.
- Added regressions for OpenAI Responses accounting, Codex WebSocket terminal usage, cost calculation, and session aggregation.
Fixes#4469
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.
- Classified OpenAI-compatible custom relays serving OpenAI model ids into the OpenAI service-tier family.
- Passed the model into OpenAI service-tier wire gating so custom relays emit service_tier when eligible.
- Reported /fast on as unavailable when the active model has no service-tier family.
Fixes#4386
Plan-approval's 'Approve and compact context' used to pass the rendered
plan-mode-compact-instructions prompt as the first positional argument
to handleCompactCommand -> session.compact(), which landed on the
session_before_compact extension hook as customInstructions. Extensions
treating that field as user focus (e.g. to bias a query-focused summary)
would then see plan-mode boilerplate instead of operator intent and
produce query-biased compactions.
Add CompactOptions.internalGuidance: a private summarizer-only channel.
session.compact() reads it into the fallback-model summarizer while the
session_before_compact hook payload still only carries the public
customInstructions arg (undefined for the plan-compact path). The
snapcompact-disable predicate and the /compact rejectsFocus guard cover
both fields so a directed summary is never silently downgraded.
Extend the interactive-mode handleCompactCommand facade + command
controller with a fourth internalGuidance parameter, and switch the
plan-approval callsite in interactive-mode.ts to route the plan prompt
through it.
Fixes#4359
Selecting "Approve and compact context" while a user turn was typed during
compaction surfaced `Failed to finalize approved plan: Agent is already
processing` and silently discarded the operator's queued turn.
`flushCompactionQueue` fires the queued user turn (fire-and-forget) before
`handleCompactCommand` returns, so by the time `#approvePlan` resumed, the
session was streaming. The previous shape aborted the queued turn and still
raced into `AgentBusyError` when `session.prompt()` ran before abort settled.
The finalize path now queues the plan-approved directive as a synthetic
follow-up when the session is streaming, and catches a racing `AgentBusyError`
from `prompt()` with the same fallback. `AgentSession.followUp()` gained a
`{ synthetic, expandPromptTemplates, attribution }` option so the hidden
execution directive lands as an agent-attributed developer message on the
follow-up queue, without flipping advisor auto-resume the user-follow-up path
does.
Fixes#4358
Idle recap crashed with `TypeError: undefined is not an object (evaluating
'H.content.filter')` when the reporter's session_stop block-decision extension
poisoned the transcript with eight persisted `session-stop-continuation`
custom entries. The primary block-reason continuation flow was intact — the
custom message correctly renders as a `role: "developer"` LLM message —
but the follow-on idle recap fired 4 minutes later, and the side-channel
provider stream handed back a `done` event whose `message.content` had been
dropped. `runEphemeralTurn`'s sanitize step at
`assistantMessage.content.filter(block => block.type !== "toolCall")` then
tripped the TypeError, propagated as "Idle recap turn failed" in the debug
log while silently muting the session.
- Normalized the provider "done" event's `message.content` to `[]` when
the shape is not an array, so a wrapper stream that drops content surfaces
as an empty recap reply instead of an unrecoverable side-channel crash.
- Guarded `#buildEphemeralSnapshot`'s in-flight-assistant preservation
branch with `Array.isArray(streaming.content)` for the same reason.
- Added a regression test that seeds the reporter's transcript shape (one
real turn, eight persisted `session-stop-continuation` customs carrying a
multi-line block reason with U+2717 glyphs) and drives a recap through a
side stream that returns `content: undefined` on `done`. The test
fails on the pre-fix tree with the exact reporter TypeError and passes
after the guard lands.
Fixes#4323
- Added `#hasPendingAsyncWake` to detect running or pending background jobs owned by the agent.
- Deferred todo reminders and `session_stop` hook passes until all agent-owned background async jobs complete.
- Ensured scheduling pauses caused by async jobs do not trigger terminal session stops or premature todo nags.
- Added `getModel` to `AgentLoopConfig` to allow runtime model resolution.
- Updated `streamAssistantResponse` to resolve the model dynamically per provider call instead of using the stale configuration snapshot.
- Enabled mid-run model switches to take effect immediately for context promotion and retry fallbacks.
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
Compared the resolved advisor runtime inputs before rebuilding on explicit enable so unchanged /advisor on calls keep the live advisor agent and history.
Added regression coverage for the unchanged-config path preserving advisor context.
Fixes#4302
Rebuilt live advisor runtimes when the explicit enable path runs so changed modelRoles.advisor values are re-resolved without restarting the session.
Added a regression test covering a live advisor moving from one role model to another through setAdvisorEnabled(true).
Fixes#4302
The settle-time reminder queues a hard 'required' tool choice paired with a
scheduled continuation. If that continuation never runs (user prompt bumps
the generation, dispose, compaction/handoff) or plan mode is exited first,
the queued directive leaked onto the next unrelated turn as a forced tool
call. Remove it by label on continuation skip, on user-initiated prompts,
and when plan mode is disabled.
Removed the live-context guard that let default model selection persist a new role without changing the active session model. The next prompt's compaction path now owns oversized-context recovery after a switch.
Fixes#4219
The branch-scan rehydrator only rebuilt `#lastCompletedRewind` and wiped
`#checkpointState` unconditionally at entry — so a branch whose latest
checkpoint had not yet been rewound came back with neither an active
checkpoint nor completed-rewind guidance. Reloading such a session (or
`switchSession()` on the same file) made the next `rewind` fail with
"No active checkpoint" even though the checkpoint entry was still the
branch leaf.
Extended the walker to also track the last unresolved checkpoint entry
and, when the branch ends without a rewind-report, seed `#checkpointState`
from that entry (id, `details.startedAt`) so `rewind` can complete
normally. Renamed the method to `#rehydrateCheckpointRewindState` to
reflect the widened responsibility and added a regression test that
truncates the branch to the checkpoint entry, resumes into a fresh
`AgentSession`, and calls `rewind` end-to-end.
Fixes#4187
Cleared checkpoint rewind runtime state when starting new sessions or creating branch sessions so stale completed-rewind guidance cannot leak into unrelated contexts.
Added regression coverage for /new and branch reset paths.
Fixes#4187
Reconstructed the completed rewind marker from the active branch so resumed sessions keep repeat-rewind recovery guidance.
Covered resume rehydration with the checkpoint rewind branch regression test.
Fixes#4187
Reconstructed the completed rewind marker from the active branch so resumed sessions keep repeat-rewind recovery guidance.
Covered resume rehydration with the checkpoint rewind branch regression test.
Fixes#4187
- Refactored the mid-run todo nudge to trigger on mutating tools (bash, eval, edit, write, ast_edit) rather than overall tool turns.
- Simplified the nudge prompt template to a concise, non-escalating reminder.
- Migrated nudge messages from "developer" role with public events to a hidden "custom" role that is excluded from the TUI and transcript.
- Introduced a separate per-cycle reminder cap of 2 to decouple mid-run hints from the user-visible stop-time escalation budget.
- Avoided triggering the todo nudge when read-only exploration tools (e.g. grep, read, glob, lsp) or errored results are returned.
Wrapped retained rewind reports with completion guidance so the post-rewind turn knows the checkpoint is closed.
Added repeat-rewind recovery errors and regression coverage for both the retained context and no-active-checkpoint path.
Fixes#4187
- Projected full tool-call arguments down to a compact summary containing only `command` and `path`.
- Truncated summarized argument fields to 200 characters to prevent inflating session log sizes.
- Replaced routine clean session disposal warnings with debug logs to reduce noise.
- Streamlined debug context in assistant message removal and agent continuation skip paths.
- Extracted duplicate user-facing compaction warning strings into a helper function.
- Thread the postmortem reason through the session teardown pipeline to the session dispose process.
- Prevent generic "dispose" logs from overwriting real triggers like SIGTERM, SIGHUP, or uncaught exceptions.
- Ensure the first teardown trigger's reason is preserved when concurrent disposal calls occur.
- Add comprehensive test coverage verifying signal-specific reason mapping inside exit diagnostics.
- Fix a minor unhandled-exception test utility expectation in input controller tests.
- Removed a duplicated branch.reverse() left by the PR #3862 merge in
SessionEntryIndex.pathTo(), which returned branches leaf-to-root and made
getLastModelChangeRole() read the oldest model change instead of the
newest — pinning the ctrl+p cycle to one slot and breaking session model
restore.
- Hardened getRoleModelCycle() to trust the recorded role only while its
resolved model still equals the active model, falling back to matching by
model after switches through alt+m, /model, or retry fallback.
- Added mutation-verified regression tests for branch ordering and the
stale-role fallback.
Cherry-pick of the reserve-budget clamp only (resolveBudgetReserveTokens + no-op compaction guard): applies compaction.ts + agent-session.ts + compaction/shake/progress-guard tests. Excludes unrelated Julia prelude timeout and ai/test churn from the PR head.
/quit and /exit hung for many seconds because AgentSession.dispose()
awaited MnemopiSessionState.dispose() unconditionally, and that path
runs consolidate() (state.ts:421) which fires a fresh LLM fact
extraction for the just-retained transcript and then awaits
flushExtractions() per owned bank. One LLM round-trip per shutdown,
no upper bound, no visible status.
- Add a timeoutMs option to MnemopiSessionState.dispose. When the cap
fires the in-flight consolidate is detached to the background and the
SQLite handles close once it settles, so writes never race a closed
handle.
- AgentSession.dispose passes SHUTDOWN_CONSOLIDATE_BUDGET_MS = 1_500 on
the user-visible shutdown path. Per-turn maybeRetainOnAgentEnd has
already retained earlier turns, so the worst case is losing episodic
promotion for the last few turns. State-replacement disposes
(mnemopiBackend.start) stay unbounded.
- InteractiveMode.shutdown surfaces a 'Closing session…' status before
dispose runs so the brief pause is explained rather than mysterious.
Two regression tests in memory-tools.test.ts cover (1) dispose returns
within the budget when flushExtractions stalls and the deferred close
still runs once consolidate settles, and (2) unbounded dispose still
runs the full #2320 consolidate-then-close pipeline.
Fixes#3641
can1357