perf(commands): load /pharn-loop and /pharn-ship quick and stop text only when needed (6.32.0)

[PrzemekGalarowicz]

Summary

/pharn-loop and /pharn-ship no longer send their --quick deltas on every request of a full run, or their stop procedure on every request before the stop. Each command keeps its entry-to-stop steps and gains two part files in .claude/commands/, each read at one named point:

• pharn-<cmd>-quick.md — read only for a --quick run, at entry, batched with a call the run already makes;
• pharn-<cmd>-close.md — the stop steps, the claims block and the Final step. The loop reads it at its first stop; ship reads it with step 7's return marker, or at an earlier STOP.

Each part has one pointer in its command. The pointer names the exact path, when the part is read, that it is trusted command text, and that it is re-read after a compaction. It also carries a not-loaded rule: stop, and never run a part from memory. No stage, order, route, check, stop decision, retry bound, ledger rule, commit rule or human gate moved. SKILLS_VERSION 6.31.2 → 6.32.0. MIN_CLI is unchanged: pharn-cli 0.7.0's own manifest and install code copy all four parts byte-identical.

Is it an optimization? (bytes measured; requests/tokens estimated)

path	sent at invocation	later loads	added requests	this text carried across the run	est. net share of run prompt (~30k / ~120k prefix)
loop full, 1 iter	77,971 → 41,104 B	close 34,340 B at the stop	+1	−33%	6.8% / 2.5%
loop full, 3 iter	same	same	+1	−38%	6.8% / 3.0%
loop full, blocked (S12)	same	same	+1	−30%	6.5% / 1.7%
loop --quick, 1 iter	same	quick 12,197 B at entry + close	+1	−17%	2.8% / 0.5%
ship full, GATE 2	67,543 → 34,104 B	close 31,042 B with verify's marker	+0	−34%	7.5% / 3.5%
ship full, STOP at regress	same	close at the STOP	+1	−34%	6.2% / 1.9%
ship --quick, GATE 2	same	quick 12,378 B at entry + close	+0	−18%	4.3% / 1.9%

Bytes are measured. Request counts and tokens are estimates from a request profile counted over the pinned steps. No live before/after run was made. An unbatched quick Read would cost about 130k tokens at the larger prefix, more than a one-iteration quick loop saves; that is why the quick reads are batched.

Tests

.dev/floor/command-family.test.mjs pins:

• the closed part set;
• each part's frontmatter, name and framing;
• one pointer per part, with its load-condition and not-loaded sentences;
• no other command text naming a part (R5b);
• no fenced line, heading or long paragraph in two files of one command;
• which step headings each file holds.

Every pin has a negative control, and the review's six mutants are kept as regressions. The existing command pins read a command together with its parts through the dev-only helper .dev/floor/command-family.mjs. Not pinned: which file a body line under an unchanged heading sits in. Advisory: that a run reads a part at its point, and that a model follows a part it received as a Read result the way it follows an invoked command body. That was not run live; the follow-ups are parts-live-run and parts-visibility-probe.

Review

An independent /pharn-dev-review found 2 floor-gate findings. Both were claims that promised more than the tests pinned, and both are fixed in d2a059d. The advisory findings are applied, or stated in BUILD.md, "After the review". Pipeline records: .dev/features/orchestrator-context/ (PLAN, GRILL, BUILD, REGRESSION, VERIFY, REVIEW, SHIP).

Local verification on the merged tree

• npm run check: every gate is green up to check:reconcile. check:reconcile is red only on the 5 files main's 6.31.2 changed; the build's epoch can't attribute the merge's writes, and CI reads NO_BASELINE.
• npm test: 4266 / 4267 pass. The one failure is 6.31.2's own macOS-only projectRoot() test, which fails the same way on origin/main.
• validate: a clean export is GREEN, 36 capabilities.
• check:changelog-entry: GREEN.

🤖 Generated with Claude Code

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: 348fb16b-bef5-4201-b551-3af7f5ceae58

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}