Animate pipeline + subagent Fable Mindset

Subagents now get the Fable mindset automatically, and with the v2 block both Opus and Sonnet close more than the 60–70% target on the reasoning habits that define the gap.

Your /goal: spin up subagents with the Fable mindset so they close 60–70%+ of the Opus/Sonnet gap to Fable, by building Nick Saraev’s animated-video workflow (video 2026-09-11) with Opus, Codex where useful, and a clean carryover into computer-use.

Why it was broken: SessionStart hooks never run inside subagents, so the mindset gate installed on 09-11 reached none of them (marker test: PREAMBLE=no). The fix is the Agent PreToolUse hook, which rewrites the prompt and sees the per-call model. The first block pointed at the full manual and 0 of 8 Sonnet agents followed the pointer; v2 inlines the loop and names the two measured habits. v2 is now the default for every non-Fable family.

gap closed = (treated − control) / (Fable − control), rows from extract-mindset’s analyzer. Observational: treated and control differ in task; small n on the treated side. The one Opus run that went past its context window fell below control, so re-injecting after compaction is the next lever for long runs.

1Codex blocker on the money gate (first item next session)

Codex’s overnight read-only review (02:37) returned REVISE: a saved quote of 0 bypasses --go even on a paid provider, a malformed spends_credits value reads as free, and the style-still API call is ungated. No credentials exist on any box, so nothing can spend today. Three small edits plus negative controls; then Codex re-review.

comms/outputs/2026-09-12-animate-pipeline/review/codex-review.log

2Mini sync deferred

Mini timed out on the tailnet and refused the key via Nano. Tarball waits at /tmp/claude-502/animate-sync.tgz on M2; re-run the tar-over-ssh when Mini is back.

3Nano hook wiring

Files are on Nano by hash. The hook itself needs Nano’s claude-file-guard to whitelist ops/reference/mindset-toggle.conf, then copy ~/.claude/hooks/pre-tool-use-agent.py.

4First real Higgsfield run

Needs an account and a credits GO. The public API (50 paths) has no Genjutsu or Wan-animate endpoint, so the transfer step is the web UI (computer-use profile, unverified) or the Codex desktop plugin the video uses.

5Optional: promote the compaction re-injection

If long subagent runs matter, add a PostCompact re-injection of the block. Not built; flagged from the 108-call data point.

The video’s workflow: recording → face still → style still (image model) → 20 s chunks → per-chunk video style transfer (Genjutsu / Wan 3.0 / self-hosted Wan 2.2 Animate) → inspect joins, retry failed gens → stitch with the original audio → silence cut (0.5 s) → frame-rate stall → export. The agent is the production manager: one batch of questions, credits quoted before spending.

• skills/animate (promoted): animate.py plan|run|inspect|finish|status, chunk.sh, stitch.sh, silence_cut.py, stall.sh, inspect_joins.sh, manifest.py; providers mock (offline, proven), higgsfield_api (style still only), higgsfield_web (task cards for a UI driver), codex_plugin; wan_local.md (Nano feasibility, no invented numbers); SKILL.md + 6 references. Three suites 22/14/26 checks passing, re-run by the orchestrator.
• skills/computer-use: profiles higgsfield-genjutsu-web.md and codex-higgsfield-plugin.md (UNVERIFIED until a real drive), references/PATTERN-auto-editing-loop.md, table rows and pointer in SKILL.md.
• Hook: ~/.claude/hooks/pre-tool-use-agent.py injects the block per family from ops/reference/mindset-toggle.conf; log ~/.claude/logs/subagent-mindset.jsonl; measurer measure/measure-gap.py with cached control and ceiling counters.
• ICM: skills-index regenerated (animate resolves at 90); blocks B105 (SessionStart never reaches subagents), B106 (marker test quoting the marker), B107 (bash guard denies docs naming credential vars), B109 (harness refuses subagent writes of REPORT*.md).