@soleur/opencode 1.0.0 → 3.354.2

package/skills/work/SKILL.md

@@ -1475,5 +1475,6 @@ See [references/key-principles.md](${CLAUDE_PLUGIN_ROOT}/skills/work/references/
 - **`VAR=$(<file) 2>/dev/null` does not suppress — bash expands before it redirects for an assignment-only command.** A simple command consisting *only* of assignments applies its redirections after the right-hand side is expanded, so a missing file still prints `No such file or directory` to the ORIGINAL stderr — and in a test harness that captures the subject's output (`cmd 2>&1`), that noise lands in the captured text every run, corrupting exactly the diagnostics you read when something fails. Brace it: `{ VAR=$(<file); } 2>/dev/null || VAR=0`. The trap is that this shape is usually introduced as an *optimization* over a correct `VAR=$(cat file 2>/dev/null || echo 0)` (dropping a fork on a hot path), so it reads as strictly better while being a regression. Same family: `x=$(cmd) 2>/dev/null` for any `cmd`. Sibling trap in the same class — a redirect that binds to the wrong thing: `VAR=$(tr a b < "$f" 2>/dev/null)` sends *`tr`'s* stderr to `/dev/null`, not the input redirection, so a missing `$f` makes the assignment non-zero and, under `set -euo pipefail`, aborts the WHOLE suite rather than failing one test — guard the read (`[[ -s "$f" ]] || VAR="<missing>"`) instead. **Why:** #6665 — both shipped into a test harness and were caught by two independent review agents. See `knowledge-base/project/learnings/2026-07-28-the-property-my-pr-existed-to-buy-was-pinned-by-nothing.md`.
 - **Emitting a forward-looking sentence as the last thing in a turn.** "Continuing to compound → ship" at the end of a response is not a handoff, it is an abandoned pipeline — the operator has to ask whether it happened. A phase-complete marker (`## Work Phase Complete`, `## Review Phase Complete`) is a CHECKPOINT: the next tool call in the SAME response must be the successor skill invocation. Stating an intention is not performing it. **Why:** #6588/PR #6716 — review completed, the turn ended on "continuing", and the operator had to prompt for the rest of the pipeline.
 - **A Monitor script must detect completion from rc/marker FILES, never `pgrep -f`.** The Bash hook blocks `pgrep -f` because it matches its own `bash -c` wrapper; a Monitor command is not covered, so a `pgrep -f "$D/run.sh" || echo GONE` arm silently never fires. And when stopping a multi-stage runner (`stage1; stage2` in one `setsid` script), kill the SCRIPT first — killing only the current `test-all.sh` lets the script advance and orphan the next stage onto `systemd --user`. **Why:** #8189.
-- **Relaunching a long-running background bash before verifying it died** - When a `run_in_background: true` task seems unresponsive, do NOT relaunch until ALL three checks confirm death: (1) broad `ps -ef \| grep -E '<substring>' \| grep -v grep` (never `pgrep -fa 'pattern$'` — anchored patterns miss processes wrapped in `doppler run -- bash ...`), (2) cache/output file size stopped growing over a 30+ second window, (3) the harness's `<task-notification>` has fired with a definitive `status` field. Log file `mtime` is NOT a liveness signal — long-running scripts buffer output between API calls. Relaunching prematurely concurrent-runs against the same API key and wastes paid spend. **Why:** PR #4156 — bench 1 was running fine the whole ~75 min, but `pgrep` with anchored pattern + stale log mtime led to two redundant bench launches (extra ~$2-3 Anthropic spend). See `knowledge-base/project/learnings/workflow-issues/2026-05-20-long-running-bench-verify-process-before-relaunch.md`. **The inverse is as costly: STOPPING one means killing the tree, not the PIDs you launched.** `run-registered-suites.sh` fans out via `xargs -P 6`, so killing its wrapper and runner leaves the workers orphaned and running — measured at 46 minutes still spawning containers, on a machine the operator had twice flagged as contended. Resolve the root's descendants (`pstree -p <root>`), confirm each `/proc/<pid>/cwd` is YOUR worktree before killing, and re-probe afterwards: a kill you did not verify is a kill you did not make. **A killed `test-all.sh` leaves its `flock -w 3600` lock waiter alive** — `kill_mine test-all.sh` matches the runner, not that child, and the orphaned `flock` holds a `lefthook` pre-commit open ("sleeping, no children") until it is killed by hand; enumerate every process whose `/proc/<pid>/cwd` is the worktree, not only the ones matching the runner's name (#7941/PR #8070).
+- **Relaunching a long-running background bash before verifying it died** - Do NOT relaunch until ALL three checks confirm death: (1) broad `ps -ef \| grep -E '<substring>' \| grep -v grep` (never `pgrep -fa 'pattern$'` — anchored patterns miss `doppler run -- bash` wrappers), (2) output size stopped growing over 30+ s, (3) the harness `<task-notification>` fired with a definitive `status`. Log `mtime` is NOT liveness — scripts buffer output. **Why:** PR #4156 — anchored `pgrep` + stale mtime launched a second bench (~$3 wasted). See `knowledge-base/project/learnings/workflow-issues/2026-05-20-long-running-bench-verify-process-before-relaunch.md`. **The inverse is as costly: STOPPING one means killing the tree, not the PIDs you launched.** `run-registered-suites.sh` fans out via `xargs -P 6`, so killing its wrapper and runner leaves the workers orphaned and running — measured 46 min still spawning containers. Resolve the root's descendants (`pstree -p <root>`), confirm each `/proc/<pid>/cwd` is YOUR worktree before killing, and re-probe afterwards — an unverified kill was never made. **A killed `test-all.sh` leaves its `flock -w 3600` lock waiter alive** — the orphaned `flock` holds a `lefthook` pre-commit open until killed; enumerate every process whose `/proc/<pid>/cwd` is the worktree, not only runner names (#7941/PR #8070).
 - **Capture `git commit`'s rc on the very next line — a trailing command owns the call's exit status.** `git commit …; git log --oneline -1` in one background call reports exit 0 when lefthook REJECTED the commit, and the log line (showing the previous commit) reads as success. Write `git commit …; echo "COMMIT_RC=$?"` and read that line, then confirm HEAD moved. **Why:** #8474 — a failed gate (`🥊 bun-test`, 471/477) went unnoticed until the next step found the changes still staged. See `knowledge-base/project/learnings/2026-09-21-every-gate-this-pr-added-failed-open-on-the-input-it-could-not-measure.md`.
+- **A diff adding or renaming members an existing derived artifact enumerates owes the FULL regen sweep — the plan's list is a hypothesis, not the set.** Run every regenerator the touched paths can invalidate (regenerate-*, sync-*-counts, *baseline* scripts) — N reds on M surfaces is ONE gap. **Why:** #10404/PR #10744 — three committed-state reds from one incomplete sweep (0400ffe62a). See `knowledge-base/project/learnings/2026-10-21-one-regen-gap-three-committed-state-reds.md`.