pkg: initial repo packaging — README, LICENSE, install.sh, skill, docs#1
Merged
Conversation
Moves the 296-commit subtree-split history into scripts/codex-fleet/ so
the canonical lib/_env.sh CODEX_FLEET_REPO_ROOT autodetect ("$DIR/../../..")
still resolves to the repo root after extraction. Adds the public-facing
packaging files that did not exist when this directory lived inside
recodee/scripts/codex-fleet/:
README.md install + usage + dependencies
LICENSE MIT
.gitignore accounts.yml + log/cache patterns
install.sh skill symlink + accounts.yml seed + env hints
skills/codex-fleet/ Claude Code orchestrator skill (lifted from
~/.claude/skills/codex-fleet/SKILL.md with
recodee absolute paths replaced by
$CODEX_FLEET_REPO_ROOT)
docs/recodee-extras.md codex-gpu-embedder + workspace colony rebuild
recipes pulled out of SKILL.md — recodee-internal,
not relevant to public consumers
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
NagyVikt
added a commit
that referenced
this pull request
May 14, 2026
) Adds the post-merge quality rail (improvement #1) and the checkpoint runner (improvement #3's data half), built around a single shared scorer primitive. The two surfaces invoke the same Python script with different inputs and different sinks — same prompt, same I/O shape, just different trigger and reader. ## The scorer (scripts/codex-fleet/lib/score-diff.py) Reads {diff, criteria, mode, pr_title} JSON on stdin, asks Claude whether the diff demonstrably satisfies each acceptance criterion in plan.md, writes {score, criteria_met, criteria_missed, reasoning} JSON on stdout. Stdlib only (urllib + json) — no pip dependency. The prompt is deliberately narrow: "did the diff satisfy these listed criteria" rather than "is this code good." That makes the call reading-comprehension rather than aesthetic judgment, which is tractable for an LLM and reasonably deterministic across runs. The exact same call shape works for "merged diff" (improvement #1) and "diff so far" (improvement #3), which is what lets one primitive serve both. ## Improvement #1: post-merge quality rail `scripts/codex-fleet/score-merged-pr.sh <PR>`: 1. gh pr view → branch, title, agent-id (parsed from agent/<owner>/...) 2. gh pr diff → full unified diff 3. find openspec/plans/<slug>/plan.md whose slug appears in the branch or title; extract its `## Acceptance Criteria` section 4. invoke lib/score-diff.py 5. merge the verdict into /tmp/claude-viz/fleet-quality-scores.json keyed by agent-id `rust/fleet-data/src/scores.rs` reads that JSON with a 5s TTL cache. `fleet::join` now takes `&ScoresFile` as a third argument and folds each agent's score into `WorkerRow.quality: Option<u8>`. `fleet-state` renders it as a third rail (Done axis: high → green, low → red) between the WEEKLY/5H rails and the STATUS chip. Un-scored agents get a same-width blank so column alignment is preserved. The score is **advisory**. It does not feed routing, claim, or autopilot decisions; the dashboard displays it for the operator to eyeball. The Python prompt scores criteria, not vibe — but it is still an LLM call, non-deterministic, and gameable, so treat low numbers as "look at this," not a fail. ## Improvement #3: checkpoint scorer (data half) `scripts/codex-fleet/score-checkpoint.sh`: - Scans `.omc/agent-worktrees/*` for active agent worktrees. - For each, computes the diff `origin/main..HEAD`, finds the matching plan.md, calls lib/score-diff.py in `mode: "checkpoint"`. - Writes /tmp/claude-viz/fleet-checkpoint-warnings.json keyed by agent-id. Intended cron cadence is ~15min. The colony task_post integration that turns a low checkpoint score into an actual blocker lives in a follow-up PR (separate repo: recodee/colony) — this lands the data half so the trigger half has something to read. ## Why one PR for #1 and #3 The user's sharpening of the plan was explicit: "Build the diff-scorer once, use it post-merge for #1, then reuse it at checkpoints for #3." Shipping them in one lane keeps the prompt + JSON schema + reader path single-sourced. A prompt tweak for one improves both; a schema break breaks both visibly at once. ## Verification cargo build --workspace: clean. cargo test -p fleet-data: 42/42 pass (was 34, +6 scores, +2 quality join wiring tests). bash -n on both wrapper scripts: clean. python3 -c "ast.parse(...)" on the scorer: clean. The actual API call cannot be exercised inside this session without an outbound network + ANTHROPIC_API_KEY — the operator runs `./scripts/codex-fleet/score-merged-pr.sh <recent-PR>` once to populate `/tmp/claude-viz/fleet-quality-scores.json` and confirm the rail lights up in fleet-state. Mock data shape verified by parses_fixture. Co-authored-by: NagyVikt <nagy.viktordp@gmail.com> Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
NagyVikt
added a commit
that referenced
this pull request
May 15, 2026
…r without --exclude-tmux (#137) Two cascading bugs in `scripts/codex-fleet/add-workers.sh` made the discovery wrapper report `only 0 healthy unallocated accounts available` and then fail with `no healthy accounts available`, even when ≥1 account in the canonical `agent-auth list` pool passed the 5h<100% / weekly<90% / not-already-active filter. Root cause #1 — cap-probe invoked with no email arguments `pick_accounts()` called `cap-probe.sh "$need"` (only the count). cap-probe's shebang is `<need_n> email1 email2 ...`: after `shift`, it iterates `for email in "$@"` over an empty list, probes nothing, and exits with 0 healthy rows. The wrapper then took the empty result as "0 healthy" and moved on. Root cause #2 — discover-accounts fail-closes when target tmux session is absent `bash discover-accounts.sh --exclude-tmux <session>` runs `tmux list-panes -s -t <session> | sed | sort | tr | sed` under `set -eo pipefail`. On a host where `<session>` doesn't exist on the default tmux server (e.g. running `add-workers.sh` outside the fleet session, or with `CODEX_FLEET_TMUX_SOCKET` unset so the wrapper degrades to the operator's default tmux), tmux exits 1, pipefail kicks in, the helper exits before reaching its python emitter, and the wrapper sees an empty tempfile. The wrapper then treated empty as "all candidates allocated" instead of "tmux filter unusable, retry without it". Fix (surgical, in-file only) 1. After the first discover-accounts call, if the tempfile is empty, retry without `--exclude-tmux`. We still keep `--exclude-active` so accounts already in `fleet-active-accounts.txt` are skipped. 2. Before invoking cap-probe, extract the email column from the discovered TSV and pass each email as a positional arg so cap-probe has something to probe. Empty discovery skips cap-probe entirely. The helper-side bug (discover-accounts.sh exiting 1 when the tmux session is missing instead of treating an empty tmux query as "no live panes to exclude") is left untouched per file-scope contract; the wrapper now compensates for it. Verified on host 2026-05-16: bash -n scripts/codex-fleet/add-workers.sh # exit 0 docker run koalaman/shellcheck:stable …add-workers.sh # only pre-existing findings bash scripts/codex-fleet/add-workers.sh 1 --dry-run # picks admin-mite (1 healthy) bash scripts/codex-fleet/add-workers.sh 2 --dry-run # picks 2 healthy Co-authored-by: NagyVikt <nagy.viktordp@gmail.com>
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.
First packaging commit for the extracted repo.
What changed
scripts/codex-fleet/so
lib/_env.sh'sCODEX_FLEET_REPO_ROOTautodetect (\$DIR/../../..)still resolves to the repo root.
README.md— install + usage + depsLICENSE— MIT.gitignore— accounts.yml + cachesinstall.sh— skill symlink + accounts.yml seed + env hintsskills/codex-fleet/SKILL.md— orchestrator skill (lifted from~/.claude/skills/codex-fleet/SKILL.md;\$HOME/Documents/recodeerefs replaced with
\$CODEX_FLEET_REPO_ROOT)docs/recodee-extras.md— codex-gpu-embedder + workspace colonyrebuild recipes pulled out of SKILL.md (recodee-internal)
Test plan
🤖 Generated with Claude Code