- Stopped and cleared the working loader before auto-compaction and auto-retry.
- Ensured stale loadingAnimation is removed so agent_start recreates the Working loader.
- Set catalog model input/output costs to 0.09/0.18 and reduced maxTokens to 65536.
- Added a smart poll-wait mode to AsyncJobManager with per-owner escalation ladder logic and a reset timer for idle pauses.
- Updated job polling to use the adaptive wait when `async.pollWaitDuration` is `smart` and to record poll completion timing for subsequent waits.
- Expanded async job settings and tests so `smart` is the default option and escalation, reset, and owner isolation behavior are covered.
- Added an `isEmpty` getter on `AgentHubOverlayComponent` to report whether no subagent rows were loaded.
- Extended `showAgentHub` with an optional `requireContent` flag so the overlay is disposed early when empty under the double-<- path.
- Added tests for double-<- gating behavior with no subagents, with subagents, and explicit hub-open behavior.
- Moved the ctrl+p model-role cycle rendering from `showStatus` to a dedicated anchored cycle container above the editor.
- Updated InteractiveMode to rebuild the cycle container in place and auto-clear the track after 4 seconds.
- Added tests that validate no scrollback stacking, in-place replacement, and timer-based clearing behavior.
- Routed `customType: "handoff"` messages to the compact divider path in Agent Hub and UI helpers.
- Added handoff summary expansion that extracts context text and strips `<handoff-context>` wrappers.
- Refactored shared divider rendering into `SummaryDividerComponent` used by compaction and handoff messages.
- Added a `snapcompact-savings.jsonl` append-only journal for per-tool savings records.
- Extended `SnapcompactInlineTransformer` to compute `savedTokens` and emit savings to an optional sink.
- Integrated `createSnapcompactSavingsRecorder` into `createAgentSession` for session-aware tool-result metrics.
- Skipped journaling entries without a session and deduped duplicate `toolCallId`s per recorder lifecycle.
- Added tests for savings-sink emissions and savings-journal read/write behavior.
- Hardened left-arrow double-tap handling with a new detector that only fires on a second tap in a 40--500ms window.
- Reused that detector for both Agent Hub opening and focused-subagent return so terminal-synthesized burst taps no longer trigger navigation.
- Added gesture tests for deliberate doubles, burst suppression, minimum-gap filtering, and focused-subagent unfocus behavior.
The continue() mock used mockResolvedValue(), which never consumed the
queued follow-up. After threshold compaction, #endInFlight ->
#drainStrandedQueuedMessages -> #scheduleQueuedMessageDrain reschedules a
zero-delay continue whenever messages remain queued, so the no-op mock spun
the drain into an unbounded microtask loop that allocated until the test
worker OOM'd (~101GB) and segfaulted, taking sibling tests down with it.
Mirror real continue() semantics (it polls and consumes the queue) by
clearing the queues in the mock, matching the sibling steer-idle-drain
tests. The drain now settles after one resume.
`restoreQueuedMessagesToEditor` prepended queued text but appended queued
images to `pendingImages`. Positional `[Image #N]` lookup at submit time
therefore broke whenever the editor draft already held pending image(s):
queued markers (numbered 1..K against their own image list) collided with
draft markers (1..M) and resolved to the wrong images; queued images
landing past slot M were orphaned.
Add `shiftImageMarkers(text, offset)` to `image-references.ts` and have
`restoreQueuedMessagesToEditor` walk each queued message in order,
shifting its markers by the running pending-image count (existing draft
images plus images already pulled in from earlier queued messages). Draft
markers stay untouched because draft images keep their original slots.
Paste markers are left alone — those are owned by the editor's paste store,
not the pending-image buffer, and queued message text never carries
unmaterialized `[Paste #N]` because the editor expands paste markers in
`getExpandedText()` before `onSubmit` fires.
Regression test seeds a draft image + a queued image-message in
`input-controller-compaction-image.test.ts` (per acceptance) and locks the
marker -> image mapping after restore. Unit tests for
`shiftImageMarkers` cover the WxH tail, Paste-marker passthrough, and
the zero-offset no-op.
Fixes#2531
- Added session-domain modules and exports for session-entries, context, listing, loader, and migrations.
- Changed persistence to async append writes plus writeTextAtomic, removing sync line APIs.
- Added compaction-aware session context rebuild with dangling tool-call cleanup.
- Added resumable session resolution with status inference, id/stem/suffix matching, and backup recovery.
Address PR review: the makeCtx session stub typed clearQueue() as
returning string[], forcing an `as unknown as` cast when the ordering
test overrode it with message objects. Type the stub to the real
RestoredQueuedMessage[] shape so the override needs no cast and shape
drift is caught at compile time.
runGuidedGoalTurn sent the rendered interview transcript to the plan/slow
provider as raw text, so a secret typed into the rough goal or an answer
bypassed the session's redaction contract. Route the transcript through the
session obfuscator before the request and deobfuscate the echoed question /
objective before it is displayed or the goal starts (no-op when no secrets
are configured).
Add a display-only `activity` field to AgentRef plus `setActivity`, fed
from the subagent progress chokepoint with a short gist of the agent's
latest intent (or current tool). Render it in the `irc list` output, the
subagent peer roster, and the TUI peer card, beside the role-derived
display name. setActivity emits no event — the roster reads on demand —
so the per-tool-call rate stays off the registry listener path. Peers
with no activity render without a dangling clause.
Refs #2470
Op: extend
When a spawner with remaining depth capacity spawns generic role-less
workers (a task/quick_task spawn without a `role`, or the same agent
cloned >=2x all without roles), TaskTool.execute appends a non-blocking
advisory steering it toward tailored specialists. Gated on DepthCapacity
so a leaf at max recursion is never nudged; the task-tool depth gate is
extracted into a shared `canSpawnAtDepth` helper reused by both the tool
gate and the advisory.
Refs #2469
Op: extend
Document the `role` parameter in the task-tool description (both the
batch and single-spawn shapes) and make tailored specialists the default
rule, not the exception. Direct a recursing worker to pass a `role` for
each sub-specialist. Activates the role field from #2467 for the model.
Refs #2468
Op: extend
Add an optional `role` field to the task spawn contract, threaded end to
end through resolveSpawnItems/spawnParamsFor into the executor. A role
injects a specialization preamble into the subagent system prompt and
becomes the subagent's display name and telemetry identity (label
normalized, length capped), so delegated trees stop being clones of one
generic worker. Empty/absent roles fall back to the agent type name.
Refs #2467
Op: extend
Addresses Copilot + Codex review on #2521:
- computeEditorMaxHeight returned 1 on terminals too small to host both the
editor and the chrome reserve, but the bordered editor never renders fewer
than 3 rows (2 border + 1 content). The cap now floors at that real minimum
(EDITOR_MIN_RENDERED_ROWS) so it no longer misreports the rows the editor
occupies; rendering is unchanged, and the contract is documented honestly
(reserve holds once terminalRows >= 7).
- #resolveOverlayLayout now always resolves maxHeight (?? availHeight), so the
maxHeight !== undefined branch in effectiveHeight, the composite slice guard,
and the number | undefined return type were dead. Tightened all three.
- Mirrored the maxHeight-default contract in the render stress oracle
(resolveExpectedOverlayLayout + compositeExpectedOverlays) so the randomized
sweep validates the real clipped geometry instead of the obsolete unclipped
one; updated the oracle helper test expectation accordingly.
Editor-height tests rewritten to assert the real contract (reserve when the
terminal can host both; pinned to the bordered minimum below that).
restoreQueuedMessagesToEditor only drained the agent steering/follow-up
queue via session.clearQueue(), but the "Alt+Up to edit" pending-bar
hint is rendered for both that queue and ctx.compactionQueuedMessages.
Messages typed while the session was compacting -- including /skill:*
follow-ups, which the follow-up path routes to the compaction queue
before its skill check -- were advertised by the hint yet unreachable,
so Alt+Up reported "No queued messages to restore".
Drain compactionQueuedMessages alongside the agent queue, merged in the
same order the pending bar renders (session-steer, compaction-steer,
session-follow-up, compaction-follow-up). The existing text-join, image
hand-back, and abort paths operate on the merged list unchanged.
Addresses Codex review on #2520:
- Cancel path: `#approvePlan` returned on `compactOutcome === "cancelled"`
without restoring the deferred pre-plan model, stranding the next turn on
the plan model and leaking `#planModePreviousModelState`. The model
transition now runs for the cancelled outcome too (the operator aborted
only the compaction, not the approval) before the early return.
- Queue-flush ordering: `executeCompaction` flushes input queued during
compaction before returning, so the post-return model switch landed after
the queued turn began streaming (deferred one turn via #pendingModelSwitch).
Added a `beforeFlush(outcome)` hook to `executeCompaction`/`handleCompactCommand`;
`#approvePlan` runs the transition through it (and idempotently re-runs it
afterward to cover the message-count short-circuit).
Tests cover the cancel restore and the before-flush ordering.
- Added new system prompts that ask models to emit titles inside `<title>` markers when forced tool calls are unavailable.
- Updated `generateTitleOnline` to use marker-based prompting and disable required `set_title` tool calls for models that do not support forced tool choice.
- Adjusted title parsing to extract the `<title>...</title>` value and fall back to stripped marker text when wrapping tags are incomplete.
- Validated queued toolChoice against active tools in agent and coding-agent sessions.
- Rejected queued forced choices with reason "unavailable" when selected tools were inactive.
- Dropped provider toolChoice payloads when requested function tools were not offered.
- Probed Tokio worker-thread support and fell back to current-thread runtime creation.
- Added a changelog entry describing that submitted user messages no longer pair OSC 133 command-start markers with missing end markers.
- Annotated the user-message component with guidance to avoid emitting OSC 133 command-start markers for submitted prompts.
executeCompaction added a transcript Spacer(1) the handoff path never adds; it leaked as an orphan blank line on cancelled/failed compaction (only the OK branch cleared it via rebuildChatFromMessages). Removed it, and on the OK branch the loader is stopped + statusContainer cleared before the rebuild so the live loader no longer composites over the reconciled transcript. The finally block stays as the idempotent cancel/fail safety net.
ToolExecutionComponent.#updateDisplay() re-ran renderResult (O(result-size)) on every invalidate - spinner ticks, stream chunks, resizes, keystrokes - so large results blocked the loop for seconds and typing lagged. It now early-returns on an unchanged dirty key (result version, expanded, partial, spinner frame, image visibility, theme epoch), and theme.ts exposes getThemeEpoch() bumped on every theme swap so cached blocks re-render on theme change.
With display.showTokenUsage on, the usage row was rendered inside the
assistant block above the turn's tool blocks. Finalizing the assistant
block was therefore deferred, and the late append recommitted the
already-committed tool rows, duplicating them in scrollback (worst with
parallel tool calls).
The assistant block now always finalizes as soon as a tool-call appears,
and the usage row is emitted as a standalone finalized block below the
turn's tool blocks across all three render paths (live event-controller,
transcript rebuild, agent-hub). The now-dead setUsageInfo/#usageInfo path
on AssistantMessageComponent is removed.
Plan approval dispatched the executor's first synthetic prompt without
checking whether the agent was still streaming the post-resolve
continuation (or a turn started by the approve-time compaction/clear),
surfacing "Failed to finalize approved plan: ... Agent is already
processing". Loop auto-submit and goal continuations hit the same throw
via submitInteractiveInput, which always called prompt/promptCustomMessage
without a streamingBehavior.
#approvePlan now aborts any in-flight turn before the synthetic prompt,
and submitInteractiveInput routes submissions through the steer/follow-up
queue (streamingBehavior: "followUp") when the session is streaming.
Non-streaming call shapes are unchanged. Extends the manual-/goal fix
(#2454) to the continuation and plan-approval paths.
Each live ToolExecutionComponent advanced its spinner glyph from its own
per-instance start time, so concurrent tool spinners showed different
frames and read as janky. Derive the glyph from a single shared monotonic
clock (sharedSpinnerFrame) so every live block animates in lockstep; the
per-instance interval still drives requestRender. The unused
#lastSpinnerAdvanceAt anchor is removed; the todo-strike counter is
untouched.
Adds a fastModeScope setting (both|openai|claude, default both). setFastMode(true)
derives the service tier from it (both->priority, openai->openai-only,
claude->claude-only) instead of hardcoding unscoped priority; the off path and
the already-on no-op are unchanged. /fast status now reports the active scope.
Default both preserves existing behavior.
Two bounded TUI overlap fixes:
- Editor max-height (coding-agent): the [6,18] clamp's floor of 6 exceeded
available space on terminals <=18 rows, letting the editor crowd the
transcript/status. Extracted a pure computeEditorMaxHeight(rows) with an
EDITOR_MIN_CHROME_ROWS=4 upper bound; identical for rows >=18.
- Overlay overflow (tui): #resolveOverlayLayout left maxHeight undefined when
the option was unset, so a tall overlay's bottom rows were dropped
off-screen. maxHeight now defaults to availHeight so every overlay is
sliced to fit and re-clamps on resize.
When a plan is approved via "Approve and compact context", #exitPlanMode
restored the pre-plan model before the compaction summarizer ran, so the
request cold-missed the plan model's warm prompt cache. The model switch is
now deferred: compaction runs on the plan model, and the switch to the
execution (slider) or pre-plan model happens only after a successful
compaction. A failed compaction stays on the plan model; cancellation is
unchanged. Also clears any queued plan-role model switch when deferring the
restore so it cannot later clobber the restored model.
can1357