workflow-fix: fable-5+xhigh agent pins autocompact-thrash subagent spawns

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix candidate raised on task #823 (emitting agent: orchestrator, /issue 823 session).

Goal

Diagnose the claude-fable-5 + effort:xhigh subagent context budget and adjust the .claude/agents/*.md model/effort pins (or spec sizes) so agent spawns do not autocompact-thrash.

Workflow gap

• Bug observed: Two consecutive planner subagent spawns (2026-07-01, task #823, /adversarial-planner Phase 1) crashed with "Autocompact is thrashing: the context refilled to the limit within 3 turns of the previous compact, 3 times in a row" — after only 9–11 tool uses, with every tool result ≤17 KB. Transcript inspection shows compaction firing every 2–3 turns from the FIRST turns, i.e. the agent's BASE context at spawn (system prompt + CLAUDE.md tree
  • the 95 KB planner.md spec + agent memory + xhigh thinking budget) is already at/near the compact threshold. The IDENTICAL brief re-run with an Agent-tool model: sonnet override completed cleanly (141K tokens, 30 tool uses, 17.5 min) — the differential implicates the fable-5 + xhigh pin combo, not the brief.
• Why it is a workflow gap: every one of the 28 files in .claude/agents/*.md pins model: claude-fable-5 + effort: xhigh; the largest specs (analyzer.md 126 KB, clean-result-critic.md 123 KB, planner.md 96 KB) are the most exposed. If the fable-5 subagent context window/thinking-budget configuration leaves ~<40 KB of usable headroom, EVERY pipeline stage (planner, analyzer, critics, implementer) can thrash and burn its spawn, stalling /issue sessions fleet-wide.
• Confidence (emitter): medium (differential evidence from one brief; other fleet sessions may or may not be hitting the same failure — check recent subagent transcripts for the thrash signature during diagnosis).

Proposed change (candidate diff sketch — refine in planning)

# Option A (per-agent effort downgrade for the largest specs):
- effort: xhigh
+ effort: high        # analyzer.md, clean-result-critic.md, planner.md first
# Option B (model re-pin if the harness exposes the long-context variant):
- model: claude-fable-5
+ model: claude-fable-5[1m]
# Option C (spec-size reduction: move on-demand sections of the >90 KB agent
# specs into .claude/rules/ referenced files)

Diagnose FIRST (measure a trivial fable-5 planner spawn's base token usage vs the compact threshold; check whether the harness honors [1m] in agent frontmatter), then pick the minimal fix. Verify by re-spawning a planner-shaped agent on the fixed pin with a real brief and confirming no compact within 20 tool uses.

Scope / surfaces

• Primary target: .claude/agents/planner.md
• Pattern hits ALL of .claude/agents/*.md (28 files, every one pins model: claude-fable-5 + effort: xhigh — verified by grep 2026-07-01); update every file the chosen fix applies to, and list them in the plan.

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• scripts/workflow_lint.py --check-asks passes; ruff on touched files passes; if workflow.yaml or CLAUDE.md change, they stay consistent with the rule file.
• This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — it MUST NOT auto-route any of its own subagents' workflow-fix candidates (recursion guard).
• NOTE for the implementing session: your own subagent spawns may hit this very bug — use Agent-tool model: sonnet overrides for your planner/critic/ implementer spawns if they thrash (the #823 workaround).

Provenance

• workflow_fix_target: .claude/agents/planner.md
• fingerprint: 24d6590ee474

Surfaced prose (orchestrator observation, /issue 823 session, 2026-07-01): two planner spawns on the agent-default model (claude-fable-5, effort xhigh) crashed with autocompact thrash (context refilled to limit within 3 turns; tool results were all <=17 KB, so base context + thinking budget is the pressure, not reads); identical brief succeeded with model=sonnet override.