fix(provider-acp): declare manual compaction support for omp

[bradhallett]

What was wrong

The provider-acp bridge already serves manual compaction — bb's compact
affordance sends a standalone builtin /compact mention, which the bridge
runs as the agent's own /compact command via session/prompt and reports
through the acp/compaction/* envelopes — but the acp-omp provider
declaration still carried the ACP tier default supportsManualCompaction: false. compactThreadContext therefore answered 409 Provider "acp-omp" does not support manual context compaction before the bridge was ever
consulted, hiding the affordance and failing plugin compactThread RPCs.
acp-opencode already overrides the same flag; omp accepts the same
/compact maintenance prompt, so the declaration was simply stale.

What changed

• plugins/provider-acp/server.ts — the acp-omp entry overrides
  supportsManualCompaction: true on top of ACP_BASE_CAPABILITIES,
  exactly matching the existing acp-opencode pattern. No new abstraction;
  the routing itself was already implemented (startCompaction in
  src/bridge.ts). Cursor, Grok Build, and Hermes Agent stay false.
• apps/server/test/public/public-thread-compaction.test.ts — new case:
  compacting an acp-omp thread returns 200 and dispatches exactly one
  turn.submit carrying createStandaloneBuiltinCompactCommandInput() with
  resumeContext.providerId "acp-omp" — the input the bridge recognizes and
  routes to startCompaction. The genuinely-unsupported provider 409
  (acp-cursor) and the active-thread 409 stay covered.
• apps/server/test/services/plugins/first-party-provider-plugins.test.ts —
  the acp-omp expectation-table entry tracks the flipped capability.
• No wire, protocol, CLI, or doc surface changes.

How you verified

• Traced the gate end-to-end at base: compactThreadContext →
  providerRegistry.supportsManualCompaction("acp-omp") → plugin
  registration's declared capabilities (supportsManualCompaction: false
  via ACP_BASE_CAPABILITIES) → 409; with the flip, the registry answers
  true and the request proceeds down the same turn.submit path the
  passing pi case exercises (the test harness registers the real
  first-party declarations, so the public test observes the actual
  declaration).
• Fail-before/pass-after executed: pnpm exec turbo run test --filter=@bb/server
  — 1923 tests, 1922 passing; the single failure
  (test/app/install-machine-script.test.ts) reproduces identically on a
  pristine main checkout at fff3ae8 (environmental, pre-existing).
  Mutation check: reverting only the supportsManualCompaction declaration
  makes exactly the two updated tests fail (the new public 200-path case and
  the first-party capability table), 1923 pass again with the flip restored.
• pnpm exec turbo run typecheck lint --filter=@bb/server --filter=bb-plugin-provider-acp
  green; pnpm exec oxfmt --check green on all three touched files.
• Bridge-side /compact routing was already covered by
  plugins/provider-acp/src/bridge/bridge.test.ts ("runs the builtin
  /compact command as compaction, not as a prompt"; failure and refusal
  paths included) — unchanged by this PR.

Fixes #2290

AGENT GENERATED

[bradhallett]

Live production evidence that the bridge compaction path works end-to-end for acp-omp — gathered today (2026-08-22) on a real thread, independent of this PR's tests.

A local auto-compaction plugin (context-compact) queues /compact for busy threads instead of calling the direct RPC. On thread thr_4tm7g4ie78 (provider acp-omp, OMP 17.4.1) the queued command drained at the turn boundary and executed a real compaction through the provider:

14:31:39Z  plugin log: auto-compact thr_4tm7g4ie78: 73111 tok (7%) → queue
           (idle acp-omp threads in the same sweep: "provider acp-omp has no
            manual compaction; skipping its threads" — the 409 gate)
14:36:50Z  turn/completed (prior turn ends)
14:36:50Z  client/turn/requested: "/compact" (builtin command mention)
14:36:50Z  turn/started → item/started contextCompaction → turn/input/accepted
14:38:17Z  thread/compacted → turn/completed

Sequence numbers on the thread's event log: 19648 (prior turn/completed), 19649–19652 (the /compact turn: requested/started/contextCompaction/input-accepted), 19656 (thread/compacted), 19657 (turn/completed). Provider session id 01a027f5-2537-7000-98be-c1612765f2f8 is stable across all four.

Two conclusions:

1. The turn-input compaction path already works for acp-omp in production — the same path this PR's new public-route test exercises (createStandaloneBuiltinCompactCommandInput → bridge → session/prompt "/compact"). The compaction above completed successfully at 14:38:17Z.
2. The 409 is purely the declaration-level gate (supportsManualCompaction: false in ACP_BASE_CAPABILITIES) blocking the idle/direct compactThread RPC. Busy threads are servable today only because the queue path never consults that gate.

So the declaration flip in this PR doesn't enable an untested path — it un-gates the direct RPC onto a path production already exercises successfully.

AGENT GENERATED

[bradhallett]

Rework addendum (reply to review of #2293)

Finding 1 (blocking — obsolete server.ts hunk): Addressed by the rebase. The branch now sits on current main (42658f987); the conflict in plugins/provider-acp/server.ts was resolved in main's favor and the one-line capability is re-expressed where #2325 put it: supportsManualCompaction: true on the acp-omp entry of plugins/provider-acp/src/known-agents.ts, exactly as the review's §6 step 1 specifies. The first-party capability table test and the acp-omp 200-path public-route case are ported onto main's layout (rebased commit 6287a8ec0).

Finding 2 (major — false thread/compacted on omp): Addressed in the same PR (new commit f8620a8ea), following the pi/#1721 precedent the review pointed at. While the compaction maintenance prompt is in flight (activePromptKind === "compaction" only), the bridge accumulates agent_message_chunk text; when the prompt ends with end_turn the completed text is classified:

• "Compaction failed: Nothing to compact (session too small)" / "Compaction failed: Already compacted" (the exact live strings from §5b) → new skipped compaction outcome → translator emits provider.warning with category: "compaction-skipped" plus a completed turn boundary and no thread/compacted — identical treatment to pi's delta translation.
• Any other "Compaction failed: …" text → turn fails with the agent's own message; no thread/compacted, no success banner.
• Anything else → unchanged completed + thread/compacted.

Detection is scoped narrowly to the compaction turn and never inspects ordinary agent traffic; the message strings are matched verbatim from the report. Covered by three new tests (two bridge-level against the fake ACP agent — which gained a FAKE_ACP_COMPACT_AGENT_MESSAGE knob to replay omp's message-then-end_turn behavior — and one delta-translation unit test), each verified fail-before/pass-after by mutation:

• early-return {status:'completed'} before the failure check → both bridge tests fail, including the exact false thread/compacted the report predicted;
• disabling the skipped→warning mapping → the translation test fails expecting the provider/warning row;
• reverting the known-agents.ts declaration → the public 200-path test fails with the 409 body and the capability table fails expected false to be true.

Finding 3 (minor — bridge behavior untested): The two bridge-level tests close this gap.

Finding 4 (minor — unrelated reformatting): Dropped during the rebase, but the repo-pinned oxfmt 0.64.0 re-introduces both hunks on main's current copy of the file (main's own blob fails oxfmt --check), so they are kept to stay format-clean; they are pure formatter output, no semantic change.

Finding 5/6 (ok): No wire/daemon/CLI changes; the skipped status is a bridge-internal envelope value (the wire still only carries existing delta kinds), so HOST_DAEMON_PROTOCOL_VERSION stays untouched.

Gates: @bb/provider-bridge-acp + bb-plugin-provider-acp test/typecheck/lint 8/8 green (258 + 76 tests); @bb/server typecheck/lint green, full suite 2005/2006 — the single failure (install-machine-script.test.ts, Abort trap: 6 in a nohup'd host-daemon join) reproduces identically with these changes stashed on clean main; scoped test/public/ + test/services/plugins/ 793/793.

AGENT GENERATED