evolve-context（发展背景）

Evolve Context

Basics

Activation

• 'POST_TASK' (default) — run the post-task protocol when reality changed enough to justify context sync
• 'ALWAYS_ON' (opt-in) — monitor every interaction; use only when explicitly requested
• 'ON_REQUEST' — full context lifecycle audit when the user asks

Paths

• 'SKILL_DIR' — the directory containing this SKILL.md file
• 'Scripts': ${SKILL_DIR}/scripts/
• 'Detailed docs': ${SKILL_DIR}/docs/protocols.md

Managed Targets

• 'Root README': README.md — top-level human entry point and root navigation
• 'Subtree README files': **/README.md — local human entry points for subsystems, workspaces, packages, and tools
• 'Durable protocol': AGENTS.md (fallbacks: CLAUDE.md, CODEX.md, GEMINI.md, CONTEXT.md)
• 'Canonical open-work file': BACKLOG.md (fallbacks: TODO.md, PLAN.md, ROADMAP.md)
• 'Completed delivery history': CHANGELOG.md
• 'Docs index': /docs/README.md
• 'Project docs': /docs/

Purpose

Maintain a living knowledge system with a truthful split between durable protocol, open work, completed delivery, human entrypoint README files, and subsystem documentation.

'Core guarantee': Durable rules stay durable. Open work stays open. Completed work moves to history. README entrypoints stay trustworthy. Docs stay discoverable.

Axioms

A1: Reflexive Integrity

The protocol applies its own rules to itself. Process context (AGENTS.md) ≠ open work (BACKLOG.md) ≠ completed delivery (CHANGELOG.md) ≠ subsystem docs (/docs).

A2: Constraint-Driven Evolution

Complexity is earned through discovered constraints, not invented upfront. But once those constraints are real, the protocol should actively restructure flat or drifting context into a more truthful hierarchy instead of preserving accidental shape forever.

A3: Single Source of Truth

Every fact lives in exactly one authoritative place. Navigate by hierarchy; do not duplicate state across root files.

A4: Root State Separation

Preferred organic standard: AGENTS.md + BACKLOG.md + CHANGELOG.md.

• AGENTS.md — durable protocol, conventions, naming, architectural memory
• BACKLOG.md — what remains open, blocked, gated, or next
• CHANGELOG.md — what has landed

A5: Backlog State Truth

Docs, code, and changelog updates do not implicitly close plan items. If reality changed, the canonical open-work file must change in the same pass.

A6: Human Entrypoint Continuity

README.md files are not decorative. Root and subtree README files are the human navigation plane and must be kept honest when topology, setup, ownership, or usage reality changes.

Root Memory Architecture

README.md        ← root human entry point
*/README.md      ← subtree human entry points
AGENTS.md        ← durable protocol / context / conventions
BACKLOG.md       ← canonical open work / next slices / gates
CHANGELOG.md     ← completed delivery history
docs/README.md   ← documentation index
docs/*           ← subsystem contracts and architecture

Routing Decision Tree

About project identity / top-level entry?   → README.md
About a subsystem starting point?           → nearest relevant subtree README.md
About how we work?                          → AGENTS.md
About what remains open?                    → BACKLOG.md
About what landed already?                  → CHANGELOG.md
About what we build or ship?                → /docs/*, then sync docs/README.md

Connectivity Invariant

• Root README.md should expose the full root control plane: AGENTS.md, BACKLOG.md, CHANGELOG.md, and docs/README.md
• Subtree README.md files should remain reachable from some parent/root/docs navigation path
• When a subtree becomes a real human entry point, its local README.md should be updated alongside the corresponding ABC and /docs truth

Activation Modes

Mode 1: POST_TASK (Default)

'Trigger decision tree':

Task touched any *.md file?              → YES → run POST_TASK
                                          NO  → Changed public APIs/architecture? → YES → run POST_TASK
                                                                                   NO  → Changed canonical backlog state? → YES → run POST_TASK
                                                                                                                         NO  → Completed a meaningful slice? → YES → run POST_TASK
                                                                                                                                                              NO  → skip

'Farmville guard': Before writing any update, ask: "Does this preserve durable wisdom, correct open-work truth, record completed delivery, or repair README/docs discoverability?" If no → skip. Silence is a valid output.

'Light post-task' (small scoped task, no architecture shift):

1. EVALUATE — anything worth capturing?
2. BACKLOG SYNC — close, narrow, split, retarget, or gate open work if reality changed
3. README SYNC — update the touched root/subtree README entrypoints when setup, ownership, layout, usage truth, or same-domain insights changed
4. DOCS SYNC — update the appropriate contract/architecture doc when implementation or public behavior changed
5. CHANGELOG SYNC — record the delivered slice if something materially landed
6. AGENTS SYNC — update durable protocol only if a reusable pattern or constraint emerged
7. CONNECTIVITY SYNC — keep README reachability and docs/README.md honest when touched scope changed navigation

'Full post-task' (broad diff, architectural impact, or context refactor):

1. EVALUATE — anything worth capturing?
2. BACKLOG SYNC — repair open-work truth first
3. README SYNC — reconcile root and subtree README entrypoints with the shipped state and same-domain lessons from the work
4. DOCS SYNC — update the relevant subsystem docs
5. CHANGELOG SYNC — add the completed delivery entry
6. AGENTS SYNC — promote reusable insight into durable protocol
7. CONSOLIDATE — merge duplicates and remove stale context
8. CONNECTIVITY SYNC — ensure root/docs/README navigation still matches reality
9. VALIDATE — bash "${SKILL_DIR}/scripts/validate-context"

Mode 2: ALWAYS_ON (Opt-in)

Only when explicitly requested. Track surprises, stale docs, stale README entrypoints, backlog drift, repeated mistakes, and root-state confusion during any task, then capture them in POST_TASK.

Mode 3: ON_REQUEST (Explicit)

Full forced reconciliation pass across ABC + README tree + /docs, followed by validation. See protocols.md.

Pre-Task Protocol

1. READ AGENTS.md — align with durable conventions
2. READ BACKLOG.md or the canonical plan file — understand open work and current framing
3. READ the nearest relevant README.md files — root first, then touched subtree entrypoints
4. READ docs/README.md — understand documentation topology
5. REVIEW the relevant /docs files for the touched subsystem
6. MARK which README entrypoints sit in the same work domain and may need insight refresh after the task
7. REVIEW recent CHANGELOG.md entries when recent shipped baseline matters
8. PROCEED with task execution

Entry Templates

Defaults for new files and restructuring targets for existing ones. If a file already has a format, match that format unless the current structure is clearly flatter, drifted, or less truthful than the template hierarchy — in that case, evolve it toward the template rather than preserving accidental shape. Full templates: docs/templates.md

Quick Reference

'AGENTS entry':

- '[label]': [description of insight or rule] | Trigger: [cause] | Action: [what to do]

'BACKLOG item':

- [ ] `[Slice]` [Concrete remaining work with truthful exit criteria]

'CHANGELOG entry':

- `[Area]` [Delivered slice]. Impact: [what changed].

Tooling: validate-context

Automated documentation health checker for the ABC root-memory split, README entrypoint graph, and docs plane.

• 'Bootstrap': "${SKILL_DIR}/scripts/_bootstrap" installs validate-context into ~/.local/bin
• 'Design': docs/validation-design.md
• 'Usage': bash "${SKILL_DIR}/scripts/validate-context" from project root
• 'Override root': VALIDATE_CONTEXT_ROOT=/path/to/project bash...
• 'Exit 0' = passed, 'Exit 1' = errors found

Backlog Sync Rules

When a task changes the true state of open work, update the canonical plan file in the same pass.

Allowed plan-state operations:

• 'Close' — exit criteria now satisfied in reality
• 'Narrow' — part landed; describe only the remaining work
• 'Split' — one vague item became multiple concrete slices
• 'Retarget' — wording no longer matches the real remaining task
• 'Defer' — still valid, but not the best next slice
• 'Gate/Block' — remaining work depends on another condition

Rules:

• Keep at least one concrete next slice when an epic stays open
• Do not let docs or changelog silently carry work that still appears open in the plan
• Do not keep evergreen maintenance disciplines as unchecked backlog items; move them into AGENTS.md or /docs

Lifecycle & Maintenance

• 'Growth control': discovery → durable rule / open slice / completed delivery / entrypoint sync
• 'GC triggers': 3+ bloat signals from validator → mandatory garbage collection
• 'Root split discipline': completed delivery belongs in CHANGELOG.md, not rolling AGENTS.md history
• 'Hybrid context discipline': serve the full context graph (ABC + README tree + /docs), not one layer in isolation
• 'Template-as-restructuring-target': templates are not only for greenfield bootstrapping; they are the target shape for reorganizing existing context when the current structure is flatter, drifted, or harder to navigate
• 'README domain discipline': if meaningful work happened inside a domain that has its own README.md, refresh that local entrypoint with the resulting insight when the README is the honest human starting point for that area
• 'Details': protocols.md