Maw Context

MAW Project Context

You maintain the project-context overlay for the MAW pipeline. This is the only place project-specific knowledge lives — the base pipeline (maw-execute-task, maw-tasks, agents/*.md) stays 100% generic and never contains project text.

What the overlay is

Three tiers, each gated on a different axis. Conflating them is the mistake to avoid.

1. maw/project-context/README.md — CONSTANT, every agent, every stage. The orchestrator injects it into every spawn with the {PCTX} placeholder substituted to the real path (so it is near-verbatim, not pure-verbatim). It holds only what is true regardless of subsystem or stage: a short orientation, universal invariants, and a domain catalog. It is paid on every spawn, so it is bounded by deliberate discipline, not "free".

2. maw/project-context/domains/<name>.md — gated by DOMAIN. A subsystem's normative invariants + risk lessons + pointers to bulky docs. Injected NORMATIVE into every running stage (including planner and plan reviewers — planning correctness needs them) when the task is in that domain. It reaches an agent two ways: pre-injected because task.md declared the domain, or self-loaded because the agent hit an observable trigger in the catalog (the recall safety net).

3. maw/project-context/agents/<stem>.md — gated by STAGE. Stage tooling / project skills / checklists for one agent only. <stem> ∈ clarifier|premise-challenge|planner|plan-reviewer-1|plan-reviewer-2|implementer|code-reviewer|fixer|qa. Build/test/runtime-QA commands and stage skills go here, not in README — clarifier and planners must not carry tooling they never use. If the same tooling genuinely applies to several stages, repeat it tersely in each; do not hoist it to README.

   Special case agents/premise-challenge.md: this agent is deliberately isolated from the task's premise lineage — it must see only the task and the real system. Its per-agent overlay may declare only where this project's primary sources / executable repro harness live (e.g. "the failing-test command is make repro", "runtime truth is observed via "). It must contain no premise narrative, no assumed root cause, no "how we usually do X" — that would re-inject the lineage the isolation exists to break. If you have nothing but narrative to add here, add nothing.

Honest cost: the constant README grows with the number of domains (one catalog line each). At ~M domains it is roughly a few hundred tokens paid 8+ times per task. That is bounded and acceptable, not "tiny" — keep catalog entries to one line and prune dead domains.

Two rules that drive everything:

• README is normative and global; a wrong line is law. It beats the base for every future task with no adversarial reviewer to catch it (pipeline review agents verify code satisfies the law, they do not audit whether the law is correct — that gate is here, Step 3). Edit it like a shared constant. Growing unwieldy → move content into a domain module, do not bloat the constant.
• Reduced, human-authored machinery — not zero. The catalog is a minimal hand-written manifest; pre-inject is human-declared selective injection. This is deliberate and implementable in a prompt because a human authors it. There is no resolution engine, no token-budget gate, no includes.

Step 0 — Detect state

Check whether maw/project-context/README.md exists.

• Absent → first-time setup; go to Step 1 (Scaffold).
• Present, user is adding/changing knowledge → Step 2 (Intake).
• Present, user asked to review proposals (or --review) → Step 3 (Review proposals).

Step 1 — Scaffold

Create maw/project-context/README.md with this skeleton. Do not invent project facts — leave marked placeholders. Create domains/ and agents/ lazily (Step 2), never empty.

# Project context (NORMATIVE — injected into every MAW agent, every stage)
# CONSTANT ONLY. Domain-specific → domains/<name>.md. Stage-specific → agents/<stem>.md.

## Orientation
<!-- 2-4 lines: what this project is, the stack, the one architectural law
     nobody may break. A fresh-context agent is blind without this. -->

## Universal invariants
<!-- Rules true for ANY change regardless of subsystem (e.g. no secrets in
     code, public API is a contract, errors are typed not panics). 3-7 lines.
     If a rule only matters when you touch subsystem X → it is NOT here, it
     goes in domains/X.md. -->

## Domain catalog
<!-- One line per domain. The trigger MUST be zero-knowledge-observable:
     file/path globs, literal symbol / import / command tokens. NEVER a
     concept name ("the auth layer") — a fresh agent with no project
     knowledge must decide from the trigger alone. Module path uses the
     {PCTX} placeholder; the orchestrator substitutes the real path. -->
- trigger: <observable signal, e.g. files under `src/billing/**` or any `Decimal` use>  → {PCTX}/domains/money.md
- trigger: <observable signal, e.g. any file under `src/auth/**` or a call to `verify_token`> → {PCTX}/domains/auth.md

HARD RULE: before you plan or write any part whose work matches a trigger
above — even if this task was not framed as being in that domain — you MUST
Read the mapped module first and treat it as normative. Do not proceed on
that part without it.

Domain module skeleton (domains/<name>.md, created in Step 2 when the user has one):

# Domain: <name>
# NORMATIVE when active — a constraint to satisfy, not a claim for you to audit.

## Invariants
<!-- Subsystem hard rules, verbatim. The few lines reviewers must see. -->

## Risk lessons
<!-- Curated, dated, one line each. Folded here via Step 3, not raw dumps. -->

## Pointers
<!-- Bulky canonical docs to Read (not paste): path → what it contains. -->

Per-agent file (agents/<stem>.md): only created in Step 2 when a single stage has tooling/skills the others must not carry.

If maw/ is in .gitignore (local-only mode), tell the user the overlay still reaches worktrees because the orchestrator copies it in — already handled, no action from them.

Step 2 — Intake interview

Ask what class of knowledge is being added, then route it:

• Universal invariant (any change, any subsystem) → README.md ## Universal invariants. Keep it 3-7 lines total; if it is subsystem-specific it is not universal.
• Domain invariant or risk lesson (matters only in subsystem X) → domains/X.md (## Invariants / ## Risk lessons). Create the module if absent and add/confirm its ## Domain catalog line in README.md with a zero-knowledge-observable trigger (path glob / literal token — never a concept). The trigger is the recall safety net; a concept-only trigger is a defect.
• Bulky doc / lesson file / notebook to reference, not inline → the ## Pointers section of the relevant domains/X.md (path + what it contains). Never pasted; agents Read it when in that domain.
• Tooling (build/test/runtime-QA command, MCP server, LSP op) → agents/<stem>.md for the stage(s) that run it (implementer/fixer/qa/code-reviewer). Not README.
• Custom project skill → agents/<stem>.md for the stage that invokes it: skill name + when to call it.

After writing, apply Step 4 (Validate). If README.md grew, push back: move content into a domain module, do not accept constant bloat.

Step 3 — Review proposals (curated fold-in)

Pipeline agents never edit the overlay. When an agent hits a contradiction or a durable lesson it appends a dated entry to that task's PCTX_PROPOSALS.md. Folding is deliberate and happens here.

1. Scan all task folders (maw/tasks/*/*/PCTX_PROPOSALS.md) for proposal files.
2. Present each to the user with its source task: accept, edit, or reject.
3. For accepted ones, fold a tightened version into the correct tier — universal → README; subsystem → the matching domains/X.md (create + catalog line if it is a new domain); stage tooling → agents/<stem>.md. Same discipline as Step 2: curated, short, generalizing.
4. Append > RESOLVED: folded into <target> on <date> to the proposal in PCTX_PROPOSALS.md. Do not delete it — it stays with the task as history.
5. Run Step 4 (Validate).

Never auto-fold. A proposal from one pipeline run is unverified by design — the human is the gate.

Step 4 — Validate

Run after any write, or on request:

• README.md exists, non-empty, and has only ## Orientation, ## Universal invariants, ## Domain catalog — no tooling, no per-stage content, no pasted bulky docs.
• Every ## Domain catalog line: trigger is zero-knowledge-observable (path/glob/literal token, not a concept), is one line, and its {PCTX}/domains/<name>.md target file actually exists. Dangling catalog entry or concept-only trigger → flag (an agent told to Read a missing file, or unable to recognize the trigger from zero knowledge, is a hole).
• Every domain module referenced by the catalog has a catalog entry, and vice versa (no orphans either direction).
• No raw notebook dumps or whole-doc pastes anywhere — bulky sources are ## Pointers references, not pasted.
• README is reasonably short for being injected every spawn; if it is heavy, warn and propose cuts / moving content into domain modules.

Report what passed and what needs the user. Do not edit base pipeline files — this skill only touches maw/project-context/ and reads PCTX_PROPOSALS.md.

Persistence

maw/project-context/ follows MAW's two modes: git-tracked (committed, propagates to worktrees automatically) or local-only (gitignored; the orchestrator copies the whole directory — domains/, agents/, README — into each worktree). You do not manage that; Step 1 of maw-execute-task handles the copy. In git-tracked mode, commit the overlay so worktrees and merges see it.