Skip to content

Agent harness distill

Audience: Maintainers + agents deciding what to stop loading Success criterion (agent-facing): Less duplicative / oral context, same judgement — gates and stems still win. Related: stems/epistemic-layout.md, improve/oral-history.md, ADR INDEX

This is not a human-docs polish exercise. The consumer is a coding agent that overweights any prose placed in always-on paths.


Success criteria (agent consumer)

Criterion Check
Map is a table, not a chat log capability-map ≪ 300 lines; no multi-KB cycle digests
One doctrine per topic Goal B: interesting-saas-context + playbook only; antagonist report ≤1 screen
ADRs are on-demand INDEX has Active set; agents not told to pre-load the ledger
Lore has a home Durable thrash classes → improve/oral-history.md once
Stamp rule is cheap Driver stamps Last-exercised cells only
Gates > prose Counter-priors / residual / dig contracts still enforce hard truths
No resurrected thrash Capability-sweep does not re-grow cycle-note archives

If an agent still needs a deleted paragraph to act correctly, the gate or stem was wrong — restore as a stem/counter-prior/gate, not as oral dump.


Done (2026-08-10)

Surface Before After
.claude/commands/improve/capability-map.md ~1700 lines (cycle digests) ~150 lines: vocab + registry + ≤5 sweep one-liners
docs/reference/antagonist-report-post-5-8.md Second full doctrine Short pointer to interesting-saas-context
docs/adr/INDEX.md Flat must-scan ledger Active table + historical full list
improve.md stamp rule Implied narrative stamps Explicit: table only; lore → oral-history
Path honesty Mixed improve/capability-map.md Canonical .claude/commands/improve/…
MCP boundary (AGENTS) Session-era freeze language Stateless protocol + CLI for long work

Not bulk-deleted (still high signal): stems, counter-priors, dig contracts, closed depth menu, strategy playbooks (action steps).


Still optional later

Candidate Why wait
HM stem length Split matrix vs judgement when editing those stems
AGENTS architecture table Drift-gated; trim only with intentional gate update
Strategy fold (gallery vs coherence) Real lanes; not pure duplication
Capability-map row count Inventory itself is useful; only prose was thrash

Anti-patterns this pass removes

  1. Capability-map as improve-log mirror — re-told every ship; used as “what to do next” instead of residual/policy/portfolio.
  2. Doctrine triple-copy — same Goal B rules in antagonist report + doctrine + playbook (agent re-decides which is canonical).
  3. ADR pre-load — 50+ long paragraphs of subsystem history as if curriculum.
  4. Session-MCP folklore as always-on fear — “MCP freezes the agent” without distinguishing long tools vs knowledge tools after stateless SDK.

Maintenance rule

When you learn a durable improve lesson:

  1. Prefer a gate or probe residual.
  2. Else one bullet in improve/oral-history.md.
  3. Else a stem / counter-prior.
  4. Never append another essay to capability-map.