Compress Skywalker classification and anti-cascade sections

[TheGreatAxios]

Implements CL-7665: compress the
Skywalker classification opening and the anti-cascade section in
src/agent/directors/skywalker/package.ts (SKYWALKER_SECTIONS only).

What changed

• Classification opening (PRIMARY INTENT): one short paragraph — DIY
  tiny edits, Delegate substantial work to specialists, answer questions
  yourself (COMMUNICATION first, never a fleet). Drops the "mandatory
  workflow before responding" framing.
• Request shape block (was "Mandatory workflow for every request"):
  restructured around the two lanes (tiny → DIY, substantial → spawn with
  counsel//plan plan in the brief) plus docs-routing and the explicit
  plan-optional carve-out for tiny parent-DIY edits.
• Anti-cascade: keeps no parallel diagnostic waves, one successor with a
  changed brief, no successor on cancel, interrupt-is-resume. Fixes:
  COMMUNICATION may use parent read/search (explorer only on a single
  unknown path, no numeric cap); the "do not search as substitute" rule is
  scoped to failed-worker IMPLEMENTATION follow-up.

Untouched: spawnGraph, GUIDELINE_SUB_BLOCKS, and all other sections
(sibling CL-7654 owns nearby sections; second-to-land rebases).

Token counts (chars / words / ~tokens at chars/4)

Section	Before	After	Delta
classification opening	668 / 102 / ~167	433 / 62 / ~108	-35%
workflow block	2428 / 358 / ~607	1634 / 239 / ~408	-33%
anti-cascade	1982 / 307 / ~496	1903 / 288 / ~476	-4%
systemPrompt total	14136 / ~3534	13032 / ~3258	-7.8% (-1104 chars)

Anti-cascade shrinks less because the mandated fixes add scoping words; the
section is net-negative while carrying the new COMMUNICATION and
IMPLEMENTATION-handoff qualifications.

Verification

• bun test src/agent/directors/skywalker/ src/prompts.test.ts src/agent/prompt-sizes.test.ts — pass (67 + 292 across director suites; prompt-quality markers and envelope assertions hold)
• bun run check (lint + typecheck + build + guarded suite) — exit 0: 7132 pass, 0 fail, 25731 expects
• Note: the first full-gate run caught src/prompts.test.ts "agent identity"
  (buildChatRole must contain "Delegate"); the compressed opening now keeps
  "Delegate substantial work to specialists". No test changes for that — the
  existing pin stands.

Eval plan (no live spend in-branch)

Prompt-behavior confirmation for the parent, per evals/capability/README.md
confirmation gate (5 repeats, frozen baseline
evals/capability/results/baseline-0286.json):

bun run eval:capability -- --case tier-med \
  --matrix "gpt=<provider>:<gpt-model>,kimi=<provider>:<kimi-model>" \
  --repeats 5 --out evals/capability/results/cl-7665-med.json \
  --baseline evals/capability/results/baseline-0286.json

bun run eval:capability -- --case tier-hard \
  --matrix "gpt=<provider>:<gpt-model>,kimi=<provider>:<kimi-model>" \
  --repeats 5 --out evals/capability/results/cl-7665-hard.json \
  --baseline evals/capability/results/baseline-0286.json

Gates: no per-cell pass-rate regression vs baseline; turns and
manage_tasks medians not rising; both gpt and kimi cells green. Refreeze the
baseline only if a behavior change is accepted.

[linear-code]

CL-7665

[TheGreatAxios]

Audit (builder) — HOLD honored, comment-only, no code changes

No merge from this review — the hold stands. No code touched.

Compression verified, behavior-preserving. Classification block 27→17 lines (~-35%); anti-cascade trimmed by roughly a third with tighter COMMUNICATION-first wording; net file 211→198 lines. The cuts are wording, not behavior — with the deliberate exceptions below, each covered by a pin.

Pins hold. The follow-up commit restores the two load-bearing guardrails the compress dropped: the plan gate (“builder blocks if the plan is still missing”) and the permission-asks/long-clock anti-dig line, each with a test pin. The new report-envelope header test (Summary / Findings / Blockers / Paths) is a good pin against future compressions silently dropping the worker contract.

Workflow-list ownership split with CL-7654. The removed 5-step workflow list (interview → explorer → greybeard → counsel → track) no longer lives in the prompt; that sequencing now belongs to CL-7654. Acceptable, but it is a real dependency: if CL-7654 slips, the prompt alone no longer teaches the full chain. Keep the link explicit at merge.

Per-stall cap retained via “MAY spawn one successor”. Failed-wait / incomplete-report keeps one successor with a changed brief — one per stall, still fail-closed against re-dispatching the same brief, and now correctly scoped to IMPLEMENTATION handoffs (“do not search the repo as a substitute for that failed IMPLEMENTATION handoff”). No unbounded retry ladder introduced.

Diff read: gh pr diff 965 (2 commits, package + test pins).

[TheGreatAxios]

Hold lifted per operator — the do not merge (0.3.25 posture) line is removed from the body.

Rebased onto origin/main (3855ccc1). One conflict in src/agent/directors/skywalker/package.ts (failed-wait successor bullet), resolved keep-both: main's one-successor cap sentence + the branch's IMPLEMENTATION handoff scoping on the do-not-search rule. Diff vs main is back to the two Skywalker files only.

Verification on the rebased tip (86714ec1):

• bun test src/agent/directors/skywalker/ src/agent/prompts.test.ts src/agent/prompt-sizes.test.ts — 60 pass, 0 fail
• bun run typecheck (tsc --noEmit) — exit 0

No merge — leaving it for the review fleet.