Project Context

Own the regenerative project-context artifacts without duplicating their workflows across harness commands. The business maps live inside their context skills as HTML (agentic-qa-core/references/business-context-maps.md); this skill generates them and updates their stale sections.

Compact Rules

• Exactly ONE mode per run: data · e2e (synonym features) · api · test-plan · refresh-all · context-skill · coverage. Load only that mode's reference; never open a second one in the same pass.
• context-skill scaffolds a project-owned <aspect>-context for an aspect the shipped context map skills (CONTEXT_MAP_SKILLS) do not cover (../agentic-qa-core/references/skill-scaffold.md §3): it cites its sources and never copies them. refresh-all never includes it. Its router row goes in the ## Project context skills table of .agents/instructions/agent-project.md, its trigger phrases in that file's triggers:; NEVER in agent-skills-and-mcps.md, which bun sync overwrites.
• coverage reads the maps, the KATA manifest and the TMS and buckets every behaviour (automated · manual-only · weak · uncovered); refresh-all never includes it. It writes only <<PRIMARY_ROOT>>/.session/project-context/coverage/ (companion files: this skill opts out of the session Phase 0) and never edits a test or a map. The weak verdict and the risk level are agent judgments with one evidence line per row; the buckets, order and diff are bun run context:coverage classify. Coverage Tech Stories are created only after a person approves the gap list, picked by weight (mkd deck for per-item choices, the question tool for one yes/no).
• Mode → reference → output: see the Mode routing table. A map mode writes ONLY its own skill's references/<map>.html; the legacy markdown files a project may hold (the skill's legacy list in CONTEXT_MAP_SKILLS, cli/lib/context-maps.ts) are read as input and never deleted. Domain vocabulary and architecture come from the maps project-discovery generates: bun run context:map business-domain-context and bun run context:map infra-context, never a .context/ file. test-plan → references/test-plan.md → the ## Master Test Plan section of the QA Master Test Plan Epic description in Jira (cached by the sync at .context/PBI/qa-artifacts/master-test-plan.md; never a local file).
• User did not name a mode → ASK. NEVER infer refresh-all from a generic "refresh the context" request.
• refresh-all runs strictly data → e2e → api → test-plan, one at a time. Each reference's own validation and approval gate must close before the next is loaded. Never skip ahead.
• Artifact missing (or a placeholder map) = CREATE mode: may write once the analysis completes. Artifact exists = UPDATE mode: generate a candidate (for a map: only its stale sections), show the diff summary, WAIT for explicit approval. NEVER overwrite an existing artifact without that approval, and NEVER regenerate a whole generated map.
• Stop the run on a hard dependency failure or a rejected overwrite. A missing SOFT dependency is not a stop: record it as a Discovery Gap and continue, exactly as the selected reference defines.
• NEVER invent business facts. Read every source the selected reference requires; anything unverified belongs under the output's mandatory discovery-gaps section, not asserted in the body.
• After a successful artifact write, add the pointer to .agents/instructions/agent-project.md ONLY when that pointer is missing. NEVER write it into AGENTS.md (boilerplate-owned, size-budgeted) or add operational prose to CLAUDE.md. Placement per framework-development/references/instructions-doctrine.md §4, then bun run instructions:check; any other instruction change is /framework-development mode instructions.
• Mode from $ARGUMENTS: when its first token matches a mode in the Mode routing table, that token IS the mode and the rest is forwarded to it unchanged (/project-context data on Claude Code, "project-context mode data" in prose on OpenCode and Codex, /skill:project-context data on Pi). No matching first token → ASK which mode.
• Before any step that uses a declared capability (metadata.requires_capabilities: db, api-schema, diagrams for the maps' figures), run the point-of-use check in agentic-qa-core/references/preflight-gate.md §8: resolve by tool-name suffix, and when no available tool provides it STOP and name the capability + how to enable it, never a silent fallback.

Read full SKILL.md when: the requested mode is ambiguous, a refresh-all chain fails mid-sequence, or you need the selected reference's own analysis steps and validation gate.

Mode routing

Resolve one mode from the invocation: the first token of $ARGUMENTS when it names a mode below (or features, the synonym of e2e), otherwise ask. Load only the reference named in that row. The former business-*-map and master-test-plan command names survive as trigger phrases only.

If the user does not identify a mode, ask which artifact to refresh. Do not infer refresh-all from a generic request.

refresh-all dependency order

Run sequentially and complete each reference's own validation and approval gate before loading the next:

1. data
2. e2e
3. api
4. test-plan

Stop on a hard dependency failure or rejected overwrite. Do not skip ahead. Missing soft dependencies remain Discovery Gaps exactly as each reference defines.

Shared contract

• Read every available source required by the selected reference. Never invent business facts.
• CREATE mode may write the missing artifact after analysis.
• UPDATE mode must generate a candidate, show the diff summary, and wait for explicit approval before overwriting. In coverage the gap list is a recomputed session file, not an artifact: the approval gate applies to the tracker write it proposes.
• Each output includes ## Discovery Gaps for unverified facts.
• After a successful artifact write, update the context pointers in .agents/instructions/agent-project.md (the project-owned overlay) only when a pointer is missing. AGENTS.md is boilerplate-owned and size-budgeted, so it never receives them. Never add operational prose to CLAUDE.md.
• A map mode writes into a business context skill that already exists (delivered by bun sync, never scaffolded here). After it writes, the mode reviews that skill's ## Rules and gotchas against the new map and PROPOSES any change; it never rewrites a rule.
• $ARGUMENTS minus the mode token are forwarded unchanged to the selected mode.