Merge main into feat/omp-adaptor (omp adapter) - CI verification - #1144
Open
Boxyboxy wants to merge 44 commits into
Open
Merge main into feat/omp-adaptor (omp adapter) - CI verification#1144Boxyboxy wants to merge 44 commits into
Boxyboxy wants to merge 44 commits into
Conversation
Add omp as a sixth verified harness adapter alongside claude, codex,
opencode, pi, and grok. omp is Pi-derived but distinct: it sets OMPCODE=1
(and CLAUDECODE=1, so detection must check OMPCODE first), runs as
process `omp`, needs --auto-approve for autonomy, uses --thinking for
effort, and exposes an active session_stop hook that forces a
continuation by returning {continue:true} - making omp a direct-blocking
turn-end harness (class with claude/codex).
- detection: fm-harness.sh prefers OMPCODE over CLAUDECODE, plus ancestry
- lock: fm-lock.sh HARNESS_RE recognizes ^omp$
- spawn: launch template (--auto-approve, -e turn-end signal ext),
--model/--thinking flags (omit max), crewmate turn_end extension
- busy detection: the busy-only "esc" hint token in fm-watch.sh and
fm-tmux-lib.sh (omp's Working... uses a unicode ellipsis)
- liveness: backends/tmux.sh classifies an omp pane as alive
- bootstrap: crew-dispatch verified()/effort_ok, secondmate-liveness set
- session-start: load-health check for the auto-discovered guard
- teardown: state/<id>.omp-ext.ts cleanup at both sites
- primary guard: tracked auto-discovered
.omp/extensions/fm-primary-turnend-guard.ts (session_stop)
- supervision: docs/supervision-protocols/omp.md + repair_line arm
- knowledge: harness-adapters skill omp section + expanded new-adapter
wiring checklist so future adapters are fully wired
- tests: new fm-omp-harness.test.sh + fm-lock.test.sh and coverage in
bootstrap, teardown, turnend-guard, supervision, liveness; tests/lib.sh
now clears ambient harness markers so detection is hermetic
Verified live on omp 16.3.12.
# Conflicts: # .agents/skills/harness-adapters/SKILL.md # README.md # bin/fm-spawn.sh # docs/turnend-guard.md
# Conflicts: # .agents/skills/harness-adapters/SKILL.md # AGENTS.md # README.md # docs/turnend-guard.md # tests/fm-teardown.test.sh # tests/lib.sh
Add omp as a sixth verified harness adapter alongside claude, codex,
opencode, pi, and grok. omp is Pi-derived but distinct: it sets OMPCODE=1
(and CLAUDECODE=1, so detection must check OMPCODE first), runs as
process `omp`, needs --auto-approve for autonomy, uses --thinking for
effort, and exposes an active session_stop hook that forces a
continuation by returning {continue:true} - making omp a direct-blocking
turn-end harness (class with claude/codex).
- detection: fm-harness.sh prefers OMPCODE over CLAUDECODE, plus ancestry
- lock: fm-lock.sh HARNESS_RE recognizes ^omp$
- spawn: launch template (--auto-approve, -e turn-end signal ext),
--model/--thinking flags (omit max), crewmate turn_end extension
- busy detection: the busy-only "esc" hint token in fm-watch.sh and
fm-tmux-lib.sh (omp's Working... uses a unicode ellipsis)
- liveness: backends/tmux.sh classifies an omp pane as alive
- bootstrap: crew-dispatch verified()/effort_ok, secondmate-liveness set
- session-start: load-health check for the auto-discovered guard
- teardown: state/<id>.omp-ext.ts cleanup at both sites
- primary guard: tracked auto-discovered
.omp/extensions/fm-primary-turnend-guard.ts (session_stop)
- supervision: docs/supervision-protocols/omp.md + repair_line arm
- knowledge: harness-adapters skill omp section + expanded new-adapter
wiring checklist so future adapters are fully wired
- tests: new fm-omp-harness.test.sh + fm-lock.test.sh and coverage in
bootstrap, teardown, turnend-guard, supervision, liveness; tests/lib.sh
now clears ambient harness markers so detection is hermetic
Verified live on omp 16.3.12.
(cherry picked from commit b2125db)
(cherry picked from commit a19a2a9)
(cherry picked from commit f0233d3)
… architecture guide (cherry picked from commit ecdd6bb)
Combines the 24-commit omp (Oh My Pi) harness-adapter branch with the 50 newer commits on main. All 14 conflicts resolved additively: every omp adapter surface is preserved and every main improvement is kept. omp support is ported into structures main added after the branch diverged: the fm-dispatch-select.sh verified-harness and effort tables, the fm_backend_tmux_agent_state five-state recovery classifier, the agent_state-based secondmate liveness sweep, the operational-input launch-brief encoder in every omp launch template, and the reworked fm-watch-arm successor contract in the omp supervision protocol.
… inside tmux) Three pre-existing main failures surfaced by running the full suite on macOS with system bash 3.2 from inside a tmux session; none were caused by the omp merge: - fm-brief.sh: assign the no-mistakes definition-of-done block with IFS= read -r -d '' instead of DOD=$(cat <<EOF ...). bash 3.2's command-substitution scanner cannot parse heredoc content holding an unbalanced quote character (the apostrophe in "firstmate's"), which made the whole script unparseable on macOS (kunchenguid#958). - fm-spawn.sh: make a failed task-metadata publication explicitly fatal. bash 3.2 does not apply errexit to a redirection failure on a compound command, so an unwritable metadata path let the spawn report success. - tests/fm-session-start.test.sh: clear TMUX for the tmux secondmate-recovery runner so the respawn under test resolves the static firstmate session instead of the host session the suite may be running inside.
Bring main's Kimi adapter, quota-aware agent-owned dispatch (drops fm-dispatch-select.sh), Claude Stop asyncRewake auto-arm, shared session-lock lib, per-harness tmux busy signatures, and the guidance/verification doc split into the omp adapter branch. Conflict resolution kept both sides: every verified-harness list now carries kimi and omp; the omp and kimi adapter sections, launch cases, teardown cleanup, and turn-end integrations coexist. omp's busy signature moved into the per-harness FM_TMUX_OMP_BUSY_REGEX_DEFAULT and its lock identity into fm-session-lock-lib.sh (FM_HARNESS_RE gains ^omp$). Fixes exposed by the merge: - fm-session-lock-lib.sh: match the command name and interpreter script basename separately instead of grepping the whole args string, so anchored short names (^pi$, ^omp$) resolve; restores pi/omp lock recognition that the lib refactor had regressed. - fm-omp-harness.test.sh: assert omp's busy signature through the new per-harness matcher plus a behavioral fm_pane_is_busy check. - docs/documentation-audiences.json: classify docs/supervision-protocols/omp.md.
…hine - fm-session-start.test.sh: the concurrent-lock test used $BASHPID, which is unset on macOS system bash 3.2, so every acquirer got an empty harness pid and none won. Derive each subshell's own pid portably via $(exec sh -c 'echo $PPID'). - fm-spawn-batch.test.sh: run_spawn left FM_CONFIG_OVERRIDE='' so config resolved to the repo's own config/, letting a captain's live config/crew-dispatch.json trip the dispatch backstop. Point it at an isolated empty config dir so the hermetic test is independent of local files.
The merge moved omp's ⟦esc⟧ into the per-harness FM_TMUX_OMP_BUSY_REGEX_DEFAULT only, but the harness-agnostic composer/submit path (fm_tmux_composer_row_state, behind fm-send's submit acknowledgement and the away-mode composer read) uses only FM_TMUX_BUSY_REGEX_DEFAULT. Without ⟦esc⟧ there, a steer to a busy omp pane on the no-box fallback could false-report a swallowed Enter. ⟦esc⟧ is omp-unique and unambiguous (unlike the claude/kimi spinners main deliberately scoped per-harness), so it is safe in the shared default; restored in fm-tmux-lib.sh and fm-watch.sh (kept byte-identical) and pinned by two omp-harness assertions. Live smoke on omp 17.1.5 re-verified: SMOKE_OK print run, busy token ⟦esc⟧ present-busy/absent-idle, pane process name omp, launch flags, OMPCODE=1, and the session_stop forced-continuation contract. Stamps updated accordingly.
Boxyboxy
marked this pull request as ready for review
July 27, 2026 18:15
Brings the branch up to date with 34 commits of firstmate main and keeps the omp (Oh My Pi) adapter working against main's current contracts. Resolved 30 conflicts against stated intent rather than by side. Verified harness enumerations became the union claude, codex, opencode, pi, pi-signed, grok, kimi, omp: main's pi-signed and the branch's omp both survive everywhere they appear. The substantive resolution is busy state. Main replaced rendered-text busy heuristics with the semantic lifecycle contract in bin/fm-busy-lib.sh, so omp now reports through that contract instead of a pane signature. omp diverges from Pi's wiring because live probes on omp 17.2.1 showed it never emits agent_settled: per turn it emits agent_start, turn_end, session_stop, agent_end, and ctx.isIdle() reads false at agent_end and flips true within 250ms. omp's per-task extension therefore takes agent_end as its idle edge and re-checks ctx.isIdle() after a short delay, keeping main's protection against auto-retries, compaction, tool loops, and queued continuations rather than dropping the guard. fm_busy_sources_for_harness gained an omp arm so those records are trusted, and fm-spawn arms the gen for omp so the extension is not dead code. The rendered-text omp token in bin/fm-tmux-lib.sh is deliberately kept: it is a delivery guard for fm-send submit acknowledgement and the away-mode pane read, not a task-state source, and removing it would leave omp the only harness without that guard. Test expectations changed deliberately, never weakened: - fm-omp-harness.test.sh: the fm-watch.sh busy-default parity assertion is dropped because main removed that default outright; every omp token assertion is kept and the test is rescoped to the delivery guard. - fm-turnend-guard, arm-pretool, cd-pretool, sessionstart-nudge: the branch's omp source-substring assertions are converted to behavioral coverage driving the real tracked extensions, following main's move away from asserting implementation-source bytes. Each replacement was mutation-verified. - fm-busy-adapter-wiring.test.sh: new omp semantic lifecycle test covering the deferred isIdle re-check in both directions. Verified live on omp 17.2.1: a real crewmate launched from the command fm-spawn builds moved busy -> idle through the real classifier, and a session_stop return-value block forced one continuation with its one-shot release intact.
Owner
|
Automated reminder: thanks for the PR! This branch currently has a merge conflict with the base branch. When you get a chance, please rebase onto (or merge) the latest base branch, resolve the conflict, and push. After that, checks will re-run and the PR will get looked at again. Noted for firstmate#1144 at |
Each rule already cost the fleet real turns, and each existed only as prose. fm-send: refuse text over 400 bytes before any send or pending-reply record. Two ~1,400-byte steers to one busy pane both reported an unconfirmed send and never arrived, while a ~230-byte pointer to the same pane landed first try. The refusal names the working pattern (append to the brief, send a pointer) and --allow-long is the explicit opt-out. The --key path is untouched. fm-brief: bake the absolute firstmate path into both the ship and the scout scaffold for any report or evidence output, and say plainly that a path inside the worktree does not survive cleanup. Three crewmates wrote evidence to a bare relative data/<id>/ path and lost it when the worktree was returned. fm-pr-merge: read the PR's check state and refuse a red PR, naming the failing checks. No configured checks and pending checks are deliberately not red; a check state that cannot be read is a refusal, because "not red" must be established rather than assumed. --allow-red-checks is the captain-authorized exception and records merge_checks_override= in the task's metadata, above the canonical pr= line so the metadata identity parse still accepts the file. Tests are colocated in the existing suites for the three scripts.
…ef refusal and write region
# Conflicts: # tests/fm-teardown.test.sh
…m one Merging origin/main brought in a corrected `tests/fixtures/*/*` family rule. The older `tests/fixtures/*` rule added on the other side survived the textual auto-merge and still matched a direct child of `tests/fixtures/`, resolving it to the literal reference "fixtures" instead of letting it fall through to `__unmapped__` as the upstream fix intends. Neither side had this defect alone; combining them created it. Removing the superseded rule restores the merged-in behavior exactly.
Stale observations were reaching the captain and workers with the confidence they had the day they were filed: decisions already shipped, retired, or explicitly out of scope were re-presented, and briefs generated from them sent workers to implement against dead premises. Three separate gaps allowed it, so each is closed in its own existing owner rather than by new machinery. Wrong authority order. AGENTS.md section 6 now states that an initiative whose plan the captain maintains outside firstmate outranks every firstmate-derived record of it, and that a hold contradicting that source is stale by definition. The rule is generic and the concrete locations stay home-local in data/, since this repo is a shared template. No verification gate on surfacing. Every ship and scout scaffold now carries a premise check: confirm the task text's central factual claims against current code, and report rather than implement when one no longer holds. It lives in the scaffold because the failure mode is firstmate not remembering to add it. decision-hold-lifecycle requires the same re-verification before a hold is relayed or briefed, which is the step it previously only required at creation. No expiry semantics. Open decisions now carry age_days, the whole days since the filing date tasks-axi already records, so an aged hold reads as a hypothesis rather than a fact. It is null when undated rather than guessed at, and the secondmate home summary passes the filing date through so cross-home decisions carry the same age.
Six omp-on-herdr defects, two behavioral and four documentation accuracy. The herdr composer classifier could not see omp's input box. omp draws its composer as a bottom-anchored rounded pair - a status row `╭── π > ... ▶───╮` directly above an input row `╰─ <typed text> ─╯` - and neither row starts and ends with the same side-border glyph, so the bordered test never matched either one and an unrelated row higher up the capture won instead. An idle omp holding unsubmitted text classified `empty`, which is the away-mode injector's go-signal, so an escalation could be typed on top of half-typed text and the concatenation submitted. Add a rounded shape that locates the bottom-most adjacent arc pair and accepts it only when herdr's native identity is omp and its status is idle, done, or blocked, mirroring the existing Pi separator-pair conjunction. Structure alone is deliberately insufficient: omp closes ordinary transcript boxes with the same `╰────╯` glyph, so shape without identity would let a leftover output box pass as a ready composer. The new branch is last in the chain, so every existing refusal keeps precedence and this can only add a recognized shape. Tests run over four raw styled captures recorded from live omp 17.2.2 panes under herdr 0.7.5, fed in verbatim because content extraction reads SGR styling. omp supervision also lost its wake path every five minutes: the watcher arm is started as an omp background async job, and omp's bash tool applies a 300s command deadline by default, so on a quiet fleet nothing woke the arm inside that window and it was killed. The watcher child survives, but the arm is the wake delivery path. Record that the arm must be started with the command deadline disabled, and say why, in the protocol, the emitted repair and ordinary-wake lines, and the away-mode daemon's own entry points. Also name omp as a native tracked background host in the away-mode guidance, add the missing omp ordinary-wake case so it no longer falls back to generic wording, correct the claim that arming reports `watcher: healthy` to its real `started`/`attached`/`FAILED` forms, and refresh the omp version stamps. Every restamped claim was re-probed on 17.2.2 rather than restamped blind: launch flags, the `--no-extensions` suppression of both `-e` and `--hook`, the lifecycle event set with its deferred `ctx.isIdle()` edge, the `session_stop` forced continuation, the `tool_call` block contract, and the process name. Facts that were not re-run are left dated at the version that verified them.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Draft for the clean-environment CI verdict only. NOT for merge (standing ruling: feat/omp-adaptor is not merged to main).
Contents: main merged into feat/omp-adaptor with all conflicts resolved, two real merge bugs fixed (session-lock lib anchored-regex matching for pi/omp; omp busy signature relocation), two pre-existing bash-3.2/hermeticity test fixes, and omp re-verified live on 17.1.5.
Expected: the 'Require no-mistakes' check fails by design (it requires a no-mistakes-originated PR body). Read the verdict from the CI workflow jobs - especially the stock macOS bash snapshot job.