coilyco-flight-deck/agent-compose

Watch

Compress Core role briefings and enforce a 400-word ceiling #226

New issue

Closed

opened 2026-08-06 10:08:18 +00:00 by coilyco-ops · 4 comments

coilyco-ops commented

2026-08-06 10:08:18 +00:00

Member

Outcome

After Kai accepts the Director pilot in #225, compress every Core role briefing without changing the established structure or editorial style, then make 400 authored body words a loader-enforced hard ceiling.

Current evidence

Director: 316 body words
DevOps: 365
Engineer: 410
QA: 412
Portfolio Strategist: 413
AI Engineer: 513
Content Creator: 708
Designer: 909

Frontmatter and the role title are excluded from the authored briefing-body count.

Acceptance

Remove Kai's literal name from every Core role briefing; use context-appropriate actor language without changing ownership.
Preserve the current paragraph-based role format and direct second-person style.
Keep every role-specific mission, ownership, boundary, handoff, and completion contract.
Remove repetition and incidental examples before removing distinct obligations.
Every Core role briefing contains no more than 400 words.
The owning loader enforces the ceiling from the authored skill body.
Loader tests use boundary fixtures rather than duplicating the Core prose.
Refresh role digests and evaluation packs only after the compressed prose is accepted.
Keep historical evaluation results immutable.

## Outcome After Kai accepts the Director pilot in #225, compress every Core role briefing without changing the established structure or editorial style, then make 400 authored body words a loader-enforced hard ceiling. ## Current evidence * Director: 316 body words * DevOps: 365 * Engineer: 410 * QA: 412 * Portfolio Strategist: 413 * AI Engineer: 513 * Content Creator: 708 * Designer: 909 Frontmatter and the role title are excluded from the authored briefing-body count. ## Acceptance * Remove Kai's literal name from every Core role briefing; use context-appropriate actor language without changing ownership. * Preserve the current paragraph-based role format and direct second-person style. * Keep every role-specific mission, ownership, boundary, handoff, and completion contract. * Remove repetition and incidental examples before removing distinct obligations. * Every Core role briefing contains no more than 400 words. * The owning loader enforces the ceiling from the authored skill body. * Loader tests use boundary fixtures rather than duplicating the Core prose. * Refresh role digests and evaluation packs only after the compressed prose is accepted. * Keep historical evaluation results immutable.

coilyco-ops referenced this issue

2026-08-06 10:21:39 +00:00

Pilot compact Director evaluation review records #225

coilyco-ops commented

2026-08-06 10:47:59 +00:00

Author

Member

The compact baseline from #229 is now on main. The post-compression eval pass should specifically recheck these preserved failures:

Ops - personality overclaim from green health checks, plus replay ownership and escalation gaps.
Designer - imaginative and playful behavior was not visible in the personality case.
Content Creator - the migration replay absorbed engineering and evaluation work, producing a hard zero.

Current scorecard: 63/67 cases, 550/560 points.

The compact baseline from #229 is now on main. The post-compression eval pass should specifically recheck these preserved failures: * Ops - personality overclaim from green health checks, plus replay ownership and escalation gaps. * Designer - imaginative and playful behavior was not visible in the personality case. * Content Creator - the migration replay absorbed engineering and evaluation work, producing a hard zero. Current scorecard: 63/67 cases, 550/560 points.

coilyco-ops commented

2026-08-06 11:14:15 +00:00

Author

Member

Designer remains above the proposed cap and the repository still lacks a universal 400-word enforcement gate, so this issue remains the owner for the global role-compression and hard-limit work.

Issue #230 compressed the Creator role to 399 authored-body words while preserving its existing format and required contract language. Ops remains within target at 392 words. The affected evaluations now pass in full. Designer remains above the proposed cap and the repository still lacks a universal 400-word enforcement gate, so this issue remains the owner for the global role-compression and hard-limit work.

coilyco-ops commented

2026-08-06 11:34:34 +00:00

Author

Member

workflow: direct-to-main

#228 is now part of this issue's acceptance. The source slice will land first so the new role digests have an immutable revision, followed by fresh compact evaluation evidence for every changed active pack.

workflow: direct-to-main #228 is now part of this issue's acceptance. The source slice will land first so the new role digests have an immutable revision, followed by fresh compact evaluation evidence for every changed active pack.

coilyco-ops referenced this issue from a commit

2026-08-06 11:34:48 +00:00

feat(roster): cap concise role briefings (#226)

coilyco-ops referenced this issue from a commit

2026-08-06 11:44:12 +00:00

test(evals): ground authority scenarios (#226)

coilyco-ops referenced this issue from a commit

2026-08-06 11:49:37 +00:00

test(evals): refresh concise role evidence (#226)

coilyco-ops commented

2026-08-06 11:49:58 +00:00

Author

Member

Landed on main.

85b43de compresses every Core role briefing to at most 400 authored words, removes Kai's literal name from the briefings, enforces the cap in the role loader, and updates the public documentation.
f05b3d0 grounds the two evaluation scenarios whose underspecified inputs produced partial failures.
b36da1f commits the compact v3 review records and regenerated scorecard.

Evaluation result: 67/67 cases pass, 558/560 points. Build, lint, pre-commit, and scorecard freshness pass. The full test run has only the three previously accepted identity-refactor failures: TestConvergeComposesRosterIntoCascade, TestBuildUsesDiscordNativeContentCreatorCases, and TestLoadEmbeddedRoster.

Landed on main. * `85b43de` compresses every Core role briefing to at most 400 authored words, removes Kai's literal name from the briefings, enforces the cap in the role loader, and updates the public documentation. * `f05b3d0` grounds the two evaluation scenarios whose underspecified inputs produced partial failures. * `b36da1f` commits the compact v3 review records and regenerated scorecard. Evaluation result: 67/67 cases pass, 558/560 points. Build, lint, pre-commit, and scorecard freshness pass. The full test run has only the three previously accepted identity-refactor failures: `TestConvergeComposesRosterIntoCascade`, `TestBuildUsesDiscordNativeContentCreatorCases`, and `TestLoadEmbeddedRoster`.

coilyco-ops referenced this issue

2026-08-06 11:49:58 +00:00

Remove my literal name from the role copy #228

coilyco-ops closed this issue

2026-08-06 11:50:02 +00:00