Skip to content

v7.205.0

Choose a tag to compare

@github-actions github-actions released this 20 Nov 04:53
7fb699d

Retention manifests: a misnamed .json file no longer disables its claim silently

What

.swarm/retention.d/ previously skipped any .json file whose name fell outside
the manifest name grammar (for example _reviewer.json, my reviewer.json), with
no error: its claimed directory was then pruned at the default 30-day horizon, and
a malformed body in such a file did not refuse the unclaimed-directory catch-all as
documented.

A non-dot .json file in retention.d whose name is not a valid manifest name now
makes the unclaimed-directory catch-all refuse to run, with the file and the reason
in the sweep result, like any other malformed manifest. This includes a .json
name followed by trailing dots or whitespace (reviewer.json , reviewer.json.),
and the check is case-insensitive (_x.JSON is refused). Dotfiles and names that do
not end in .json (editor scratch files, x.json.bak) are still ignored.

Behavior change: only the catch-all is refused

A malformed or misnamed manifest now skips only the tier-3 unclaimed-directory
catch-all (its retention.dry_run listing included). Every directory-family pass
and the other retention passes still run, because they never read the claims. This
narrows the v7.204.10 behavior, where an unusable manifest refused the whole sweep
and stopped family prunes as well. The catch-all stays fail-closed: nothing under
.swarm/ that is unclaimed is deleted while a claim cannot be verified.

The reason is visible only in the debug log (OPENCODE_SWARM_DEBUG=1), because
neither production caller reads the sweep result. A grammar-valid stray .json
(for example notes.json or owner-copy.json) already refused this way before
this change; the change widens the set of names that do. OS file-manager duplicates
such as owner (1).json or Copy of owner.json used to be skipped harmlessly and
now stop the catch-all until renamed or removed.

Migration

If a project has a hand-made retention.d/_x.json (or a name with a space or
non-ASCII letter, or a file-manager duplicate), rename it to start with a letter or
digit and use only letters, digits, ., _ and -, or delete it if it is not a
claim. Until then the catch-all is skipped, the file is named in the debug log
(OPENCODE_SWARM_DEBUG=1), and no unclaimed directory is deleted.

Doc corrections to the v7.204.10 release notes (#3727), which stay as published:
the tier-3 rule is "no future-dated file" (age is measured from regular files),
not "no future-dated entries", and the retention result is surfaced only in the
debug log because neither production caller reads it.


fix(delegation-gate): the completion gate no longer wedges OpenCode v2 Code Mode sessions

  • On an OpenCode v2 host every plugin tool is reached through the host-owned
    Code Mode execute wrapper ({ code }), and each nested tools.<name>()
    call re-enters tool.execute.before under its own name. The completion gate
    judged the wrapper itself as an opaque, non-read-only tool, so once a task
    reached tests_run it denied every plugin call — including the
    update_task_status remedy its own message prescribes and the exempt
    check_gate_status — and the task could never leave the gate. The PR author
    reported this on a live session (11 consecutive
    TASK_COMPLETION_GATE_VIOLATION denials for one task over about 11 hours);
    that observation is not reproducible from the repository.
  • The wrapper call that the v2 adapter marked (execute with a string code
    program) now passes the completion gate. Each nested tools.<name>() call is
    still gated under its own name, so hooked mutating tools remain blocked while
    a task awaits completion. An execute call the adapter did not mark stays
    gated. Limits: the mark is not proof of provenance (a plugin or MCP tool that
    is also named execute and takes a string code is marked too), and the gate
    does not inspect the program, so sandbox effects no hook sees (for example
    fetch) are not gated by this check. Unlike Full-Auto, this exemption needs
    no permission_policy.trust_code_mode_wrapper opt-in; Full-Auto still
    classifies the wrapper separately.
  • The violation message now also names the audited escape for a task that
    cannot complete: update_task_status with status="blocked" (recorded in
    the plan ledger and task evidence) releases the gate. Reopening a blocked
    task takes the audited blocked to in_progress repair, which requires
    force: true with the exact CAS and audit fields (see the recovery runbook).
    /swarm recover and /swarm reset-session do not release the gate: the
    awaiting state lives in task evidence and plan.json, not in session state.

Keenable is now an optional General Council web search provider

What

council.general.searchProvider accepts a third value, "keenable", alongside
"tavily" (still the default) and "brave". Nothing changes unless you set it.

  • Keyless by default. Keenable works without an API key: the council
    pre-search pass posts to https://api.keenable.ai/v1/search/public, so General
    Council web search can run before a Tavily or Brave key exists.
  • Optional key. If KEENABLE_API_KEY is set, requests go to the keyed
    endpoint https://api.keenable.ai/v1/search with an X-API-Key header.
  • The shared searchApiKey is never sent to Keenable. A council.general.searchApiKey
    value left in your configuration for Tavily or Brave is intentionally ignored when
    the provider is keenable, so a project-level config that only switches
    searchProvider cannot forward a key you set for another vendor. Keenable
    reads its key from the KEENABLE_API_KEY environment variable only.
  • Same safety rails as the other providers. Requests share the existing timeout
    wrapper and typed errors, and refuse HTTP redirects (redirect: 'error').
    Results map onto the existing WebSearchResult shape (title, url, and
    snippet with a description fallback).
  • Wiring. The council and deep-research skills, the architect prompt, the
    web_search tool description and the command help no longer say a Tavily or
    Brave key is required, so a keyless Keenable configuration is no longer refused
    before the first search. The JSON schema and docs/configuration.md list the
    new value.

Why

Tavily and Brave both need a key before the council can search the web at all.
Keenable removes that setup step for first-time users who want grounded council
answers.

Migration

No action required. To opt in, set council.general.searchProvider to
"keenable"; add KEENABLE_API_KEY to your environment only if you have a
Keenable account.

Caveats

  • Privacy. Keyless requests are sent to Keenable's public endpoint, which logs the
    query text and your IP address. Keyless limits are 1,000 requests per hour and
    10 per second per IP (see https://docs.keenable.ai/rate-limits). Choose
    Tavily or Brave, or set KEENABLE_API_KEY, if that is not acceptable for your
    queries.
  • Freshness is not mapped. The provider does not send a time-range filter to
    Keenable, so the web_search tool does not report freshness as applied for
    this provider. Stale years are still stripped from the query text.
  • Tavily and Brave behavior is unchanged.