workflow-fix: trim heavyweight agent specs + size lint (subagent context headroom)

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from an orchestrator observation on task #825 (emitting agent: orchestrator, /issue 825 session, 2026-07-02).

Goal

Restore working context headroom for heavyweight subagent spawns by trimming the largest agent specs (move recipe detail to on-demand rules) and adding a size lint.

Workflow gap

• Bug observed: five subagent deaths in one /issue 825 session: 3× planner spawns died to autocompact thrash (context refilled to limit within 3 turns, 3× in a row; total_tokens reported 0), 1× experiment-implementer died with "Prompt is too long" after 5 tool uses, 1× codex-critic composer died to the same thrash. Lighter-spec agents (critic 65KB — borderline, general-purpose ~0KB, implementer 16KB) completed fine at 150-220K total tokens.
• Why it is a workflow gap: agent base context stacks: agent spec (planner.md 95,615 B; experiment-implementer.md 71,578 B; critic.md 65,514 B) + always-on agent MEMORY.md + project CLAUDE.md (130,424 B) + its rule imports + user-global CLAUDE.md/SOUL/USER + MCP tool schemas. The heavyweight specs leave so little headroom that a handful of ordinary tool results (an arXiv search, a windowed file read) trips permanent autocompact thrash — the agent cannot complete ANY bounded research task. The orchestrator had to compose the plan itself (permitted by the refusal-recovery recipe but the wrong default) and substitute implementer for experiment-implementer.
• Confidence (emitter): high (reproduced 5×, with a clean size gradient across agent types).

Proposed change (candidate diff sketch — refine in planning)

• Audit .claude/agents/*.md by size; for planner.md / experiment-implementer.md / critic.md move per-scenario recipe blocks (marker-training walkthroughs, incident post-mortems, worked examples) into .claude/rules/*.md with paths: frontmatter or short pointers, targeting ≤35 KB per spec.
• Add workflow_lint.py --check-agent-spec-size (WARN >40 KB, FAIL >70 KB) so regrowth is caught at commit time.
• Consider trimming always-loaded agent MEMORY.md indexes to the 150-char-per-line contract.

Scope / surfaces

• Primary target: .claude/agents/planner.md, .claude/agents/experiment-implementer.md, .claude/agents/critic.md, scripts/workflow_lint.py
• Grep the workflow surface before editing; list every hit in the plan.

Constraints / invariants

• Workflow-surface only. scripts/workflow_lint.py --check-asks passes; ruff passes on touched files.
• No behavioral contract changes to the agents — content relocation + linting only (anything more is architectural and parks at plan_pending).
• This session runs under the workflow-fix recursion guard once spawned.

Provenance

• workflow_fix_target: .claude/agents/planner.md, .claude/agents/experiment-implementer.md, .claude/agents/critic.md, scripts/workflow_lint.py
• fingerprint: (computed at filing)

Orchestrator observation (verbatim summary): "planner spawns for #825 died 3× to autocompact thrash before any plan text was produced; experiment-implementer died 'Prompt is too long' after 5 tool uses; the codex-critic statistics composer thrashed. Agent-spec size is the controlling variable: 95KB/72KB specs died, 16KB/0KB specs succeeded on the same task in the same session."