workflow-fix: planner/critic subagent spawns still autocompact-thrash post-#829

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix candidate raised on task #834 (emitting agent: /issue orchestrator).

Goal

Diagnose the spawn-time context composition for the planner / critic agent types and shed enough always-on weight that a fresh spawn can read a ~450-line plan plus a ~650-line module without autocompact thrashing.

Workflow gap

• Bug observed: planner- and critic-type subagent spawns autocompact-thrash ("context refilled to the limit within 3 turns of the previous compact, 3 times in a row") within 4-9 tool uses, even after the #829 spec trim (planner.md 95,759→63,888 B). 5 of 6 such spawns failed while driving task #834 on 2026-07-02: 2× planner (plan drafting), 1× planner-as-fact-checker, 2× critic (statistics + alternatives lenses). The SAME briefs re-run as general-purpose agents (16 KB-class spec) succeeded at ~160-170K total tokens each, so the work itself fits — the agent-type baseline does not.
• Why it is a workflow gap: the per-agent-type context composition (spec + memory frontmatter + always-on CLAUDE.md/import pile + whatever else the harness pins at spawn) leaves too little compactable headroom for the role's own mandated reads; #829's trim was insufficient for these two roles. Every /issue pipeline stage that spawns them (adversarial-planner Phases 1, 1.5, 2) silently degrades to crashed rounds + orchestrator fallbacks.
• Confidence (emitter): medium (symptom high-confidence + reproducible; root-cause attribution between spec size, memory loading, and harness pin composition not yet isolated).

Proposed change (candidate diff sketch — refine in planning)

• Measure actual spawn-time baseline tokens for planner / critic / general-purpose (e.g. a probe agent that reports its own context usage on turn 1) to isolate the dominant term.
• Trim .claude/agents/planner.md (63.9 KB) and .claude/agents/critic.md (65.5 KB) further: move worked examples / incident post-mortems / per-lens long-form guidance into on-demand reference files loaded by pointer, keeping the always-on role spec ≤ ~20 KB (the empirically-working general-purpose class).
• Tighten the #829 size lint threshold to whatever the probe shows is safe.

Scope / surfaces

• Primary target: .claude/agents/planner.md, .claude/agents/critic.md
• Secondary: scripts/workflow_lint.py (size-lint threshold from #829); consider the same treatment for other >60 KB specs (analyzer.md 112 KB, clean-result-critic.md 104 KB, code-reviewer.md 69 KB) if the probe confirms spec size is the dominant term.

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• scripts/workflow_lint.py --check-asks passes; ruff on touched files passes.
• No role-content loss: moved material stays reachable via explicit pointers.
• 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).

Provenance

• workflow_fix_target: .claude/agents/planner.md, .claude/agents/critic.md
• fingerprint: 23e7e522d498

target_file: .claude/agents/planner.md, .claude/agents/critic.md bug_observed: planner- and critic-type subagent spawns autocompact-thrash within 4-9 tool uses even after the #829 spec trim (5 of 6 spawns failed on task #834, 2026-07-02) why_workflow_gap: the agent-type spawn-time context composition leaves too little compactable headroom for the role's own mandated reads; #829's trim was insufficient for these two roles proposed_change: diagnose spawn-time context composition for planner/critic agent types and shed enough always-on weight that a fresh spawn can read a 450-line plan plus a 650-line module without autocompact thrashing diff_sketch: |

• probe spawn-time baseline tokens per agent type
• move worked examples / incident post-mortems out of planner.md + critic.md into on-demand reference files (target ≤ ~20 KB always-on)
• tighten the #829 workflow_lint size threshold accordingly confidence: medium related_task: #834