* fix(channels): time-gate progress drafts and defer narration until visible
Progress drafts started immediately on the second work event, so the 5s
initial delay never filtered small fast tasks and Discord showed narrated
status for turns that finish in seconds. The gate is now purely time-based
(explicit startNow paths like approvals/patches are unchanged), and the
progress narrator no longer spends utility-model completions before the
channel draft is visible - notes buffer so the first visible narration has
full context.
Surface: src/channels/streaming.ts gate, src/auto-reply/reply/progress-narrator.ts,
Discord draft-preview wiring; test updates across discord/matrix/msteams/slack.
Refs #105872
* feat(channels): model-preamble status headline for progress drafts
In progress mode the draft's status line is now the model's own typed pre-tool
preamble whenever it is fresh (always on, no new config); utility-model
narration fills the slot only when the model has been quiet for ~20s or emits
no preamble. Preambles never start the draft early - the time gate still owns
the small-vs-big decision - and narration generation is suppressed while a
fresh preamble exists, saving utility calls. Discord and Telegram render the
headline; the opt-in streaming.progress.commentary lane and its receipt
counters are unchanged.
Refs #105872
* refactor(channels): split status-text and telegram preview helpers to honor the LOC ratchet
The LOC ratchet forbids raising baseline entries, so the headline feature's
growth is paid for with clean splits: reasoning/commentary text normalization
moves to src/channels/progress-draft-status-text.ts, Telegram progress preview
rendering to extensions/telegram/src/progress-draft-preview.ts, and small
ownership moves (formatCompactionModelRef, resolveFollowupAbortSignal, shared
CLI preamble gate, Discord preamble item-event handling) shrink the runners
back under their pins. Baseline entries lowered to exact new counts; the
compositor entry is removed (424 lines, under the 500 ceiling).
Refs #105872
* test(auto-reply): add resolveFollowupAbortSignal to the queue mock factory
The explicit ./queue.js vi.mock factory must export every binding prod touches;
the relocated resolveFollowupAbortSignal was missing and failed the whole
followup-runner suite.
Refs #105872
* fix(channels): sanitize preamble headlines and retry skipped narration
Review findings: preamble headlines now strip inline delivery directives and
reject silent-control tokens (NO_REPLY) before rendering, sharing the
commentary sanitizer; rejected input keeps the previous valid headline. The
narrator schedules cancellation-aware retries when a run is skipped for draft
invisibility or preamble freshness, so single long tool calls still narrate
and the ~20s utility filler fires without needing another event. Silent
preambles no longer stamp freshness or add narration notes.
Refs #105872
* fix(channels): bound narrator visibility polling and gate Telegram headline mode on accepted preambles
Visibility retries stop after 30 one-second attempts (turn completion has no
narrator disposal signal) and re-arm on new activity; Telegram derives its
headline-rendering state from the compositor's sanitized headline instead of
raw payload text, so directive-only or NO_REPLY preambles keep structured
tool-line rendering. Docs scope the preamble headline to Discord and Telegram.
Refs #105872
* fix(telegram): read headline state from the compositor at render time
The progressHeadlineActive mirror flag was set after pushPreambleHeadline had
already rendered, so the first valid preamble arriving after the draft opened
rendered through the structured-line path and the headline never appeared
until the next event. The renderer and formatLine now read the compositor's
sanitized hasStatusHeadline directly and the mirror flag is deleted; adds an
ordering regression test. LOC baseline lowered to the exact new count.
Refs #105872
* fix(channels): let the opt-in commentary lane own preambles and scope narration docs to Discord
With streaming.progress.commentary enabled, the status headline would replace
the documented interleaved 💬 lines for any turn with a preamble (the exact
event the lane consumes). The compositor headline now declines when the
commentary lane is enabled, preserving that lane's shipped shape, and docs
state the utility-model filler is Discord-only today (Telegram's headline is
preamble-only; it has no narrator wiring).
Refs #105872
* fix(msteams): cancel the pending progress gate at finalize
With the purely time-based gate, every fast Teams progress turn leaves a
pending 5s start timer; the SDK resets its stream id on close, so a late
timer fired renderInformativeUpdate against the closed stream and posted a
stale working card below the final answer. finalize() now cancels the gate
and the informative renderer refuses to run once final text is queued.
Refs #105872
* docs(channels): record the headline-vs-commentary ownership decision inline
Refs #105872
* style: oxfmt the late review-fix rounds
Four files from the final review-fix commits were never oxfmt-formatted and
failed the shared CI formatting step; baseline lowered for the one-line
telegram shrink.
Refs #105872
* fix(channels): harden progress draft lifecycle
* fix(channels): refresh repeated preamble items
* fix(plugin-sdk): expose progress lifecycle options
* style(auto-reply): simplify narrator reset
* docs(channels): refresh progress draft docs map
* fix(auto-reply): clear queued narrator request context
* fix(discord): cancel delayed progress gate at final
* docs: refresh plugin sdk api baseline
* feat(agents): derive a provider-declared default utility model when unset
When agents.defaults.utilityModel is not set, utility tasks (titles, progress
narration) now use the primary provider's declared small model
(modelCatalog.providers.<id>.defaultUtilityModel: OpenAI -> gpt-5.6-luna,
Anthropic -> claude-haiku-4-5). Auth is inherent because derivation follows
the agent's primary provider. Setting utilityModel to an empty string
disables utility routing entirely; narration turns on automatically when a
default resolves and stays off otherwise.
Formatting/lint/tests verified on Testbox (oxfmt, oxlint, plugins:inventory,
docs:map, import-cycles, check:test-types, 4 test files).
* fix(ai): drop the temperature parameter for models that reject it
The GPT-5.6 family 400s on temperature via the Responses API (live-verified:
gpt-5.6-luna/-terra reject it; gpt-5.5 and gpt-5.4-mini/nano accept it).
supportsOpenAITemperature gates all three OpenAI payload builders, with a
catalog compat override (compat.supportsTemperature) declared for the 5.6
family in the openai manifest. Without this, utility tasks (titles,
narration) would fail once gpt-5.6-luna becomes the derived default.
Gates on Testbox: oxfmt, oxlint, plugins:inventory, import-cycles,
check:test-types, 6 test files. Live proof on Testbox: gpt-5.6-luna and
claude-haiku-4-5 one-shot completions succeed with temperature requested.
* fix(agents): carry the primary model's auth profile onto the derived utility default
A primary like openai/gpt-5.5@work previously reached utility tasks via the
profiled fallback; the derived default shares the provider, so its trailing
auth profile carries over instead of silently switching to default
credentials (Codex review). Gates on Testbox: oxfmt, check:test-types, tests.
* docs: realign the manifest provider-fields table after adding defaultUtilityModel
* feat(channels): narrated progress drafts + activity receipt on the final answer
Progress mode replaces raw tool lines with short utility-model narration of
what the agent is doing (streaming.progress.narration, default on, requires
an explicit utilityModel). On Discord the final answer now carries the
-# activity receipt and the working draft is deleted once the answer lands,
so busy channels keep no orphaned tool log above the reply.
Formatting verified remotely via Testbox oxfmt --check (hook bypassed:
no node_modules in this worktree).
* fix(channels): keep narration toggle independent of channel-default stream mode
Discord resolves its own progress default, so the resolver must not re-derive
mode with the generic partial fallback (narration was off for unset config).
* fix(auto-reply): honor status-only command text in narration model input
streaming.progress.commandText: "status" hides raw exec/bash text from the
channel draft; narration input now mirrors that policy so the utility model
never receives more command detail than the draft shows (Codex review P2).
Formatting verified remotely via Testbox oxfmt --check.
* fix(auto-reply): share the draft's command-tool set for narration and regen channel metadata
Reuse isCommandToolName (exec|shell|bash) so narration's commandText policy
matches the draft formatter exactly, and regenerate bundled channel config
metadata for the new streaming.progress.narration key (Codex review round 2).
Gates verified on Testbox: config:channels:check, oxfmt --check, 4 test shards.
* fix(channels): clear stale narration when the narrator stops mid-turn
An empty narration update now falls the draft back to raw tool lines, and the
narrator emits that clear when it disables after consecutive failures or the
per-turn cap, so drafts never pin stale status text (Codex review round 3).
Gates verified on Testbox: oxfmt --check + 4 test shards.
* chore(config): regen bundled channel metadata after rebase onto main
* chore: CI fixups — lint nits, test harness types, docs map, SDK surface budget
Two deliberate public SDK additions (resolveChannelStreamingProgressNarration,
isCommandToolName via the streaming wildcard re-export) bump the pinned
public-surface budgets to current counts (exports 10488, callable 5235).
Gates verified on Testbox: oxfmt, targeted oxlint, docs:map:check,
config:channels:check, check:test-types, 5 test shards.
Add a general typed tool-progress contract so long-running non-exec tools can emit public channel progress without overloading model-facing tool content.
`web_fetch` now uses the generic delayed progress helper: it shows `Fetching page content...` only when the fetch is still pending after five seconds, clears the timer on completion/abort, passes the abort signal into guarded fetch, and avoids provider fallback or cached success after cancellation. The subscriber path accepts only explicit `visibility: "channel"` and `privacy: "public"` progress metadata, while untyped tool partials and exec output keep their existing behavior.
Docs now explain typed progress, delayed producer examples, and the `web_fetch` timing behavior.
Proof: `pnpm test src/agents/tools/web-tools.fetch.test.ts src/agents/embedded-agent-subscribe.handlers.tools.test.ts -- --run`; `pnpm docs:check-mdx`; changed-file `pnpm exec oxlint ...`; `git diff --check`; autoreview clean.
