Context Maintain

Context Maintain

Use this after init and catch-up to update context, capture lessons, stress-test plans, maintain PLAN.md, close NOW.md, run automatic Dream/Compact consideration, or reflect after corrections/failures.

Trigger

Use this when structure, workflow, dependencies, rules, progress, findings, decisions, or closeout state changed; the user teaches durable context; repeated correction/failure happens; tool output contradicts expectations; you ignored context or instructions; the user says "remember", "always", "never", "use X", or "that's wrong"; or you are ending and need to update NOW.md.

For ordinary uncertainty before the first attempt, inspect or ask instead.

Maintain Loop

1. Read NOW.md, then concise CONTEXT.md as the always-read project layer.
2. Run node scripts/context-index.js hydrate "update context after current work" or a task-specific maintenance query before opening PLAN.md, chunks, or bulky/task-specific context.
3. Use selected cards to choose only the target source sections needed for edits.
4. Classify information as durable context or task-local state.
5. Update the smallest appropriate source section: task-local findings/progress/ decisions to PLAN.md; durable concise terms, rules, invariants, or lessons to CONTEXT.md; resume packet to NOW.md last.
6. After updating CONTEXT.md, PLAN.md, or NOW.md, always run node scripts/context-index.js update.
7. Prune stale or duplicate entries instead of appending forever.
8. Run the project's verification command when the change affects behavior.
9. Run the Should Dream check before closeout or final response.

Confidence

Only high and repeated medium signals are durable-memory candidates.

Plan Stress-Test

Use when a plan, taxonomy, workflow, or context model needs pressure-testing before implementation.

1. Read NOW.md, then CONTEXT.md, and PLAN.md if it exists.
2. Check vague terms against CONTEXT.md ## Language.
3. Inspect files, docs, tests, and cheap dry runs before asking questions.
4. Ask only where human judgment changes the agent's direction.
5. If you ask, ask one question at a time and include a recommended answer.
6. Route task-local uncertainty to PLAN.md; route durable terms, invariants, and workflow constraints to CONTEXT.md.
7. Name the next concrete checkpoint and verification command.

Reflect Mode

1. Stop the automatic retry.
2. Reconstruct the trace from files, tests, tools, or the correction.
3. Name the bad assumption in one sentence.
4. Classify the signal: explicit rule, direct correction, repeated pattern, strong approval, or low-confidence observation.
5. Route the lesson.
6. Choose the smallest correction.
7. State the checkpoint that will prove the next attempt worked.
8. Resume with the correction.

Dream/Compact Mode

Dream is an automatic semantic consolidation check inside every context-maintain run. It decides what future agents should retrieve, at what layer, not just what can be shortened.

Should Dream Check

Dream when future catch-up would materially improve: completed or redirected task, bulky NOW.md, stale/duplicated PLAN.md, obsolete blocker, repeated correction, duplicated fact, large CONTEXT.md section, or lengthy durable detail that should not be always loaded. If no condition is true, skip Dream silently. Do not log skipped checks.

Dream Pass

When Dream is useful, edit directly; do not ask for approval.

1. Detect drift with node scripts/context-index.js stats plus NOW.md, PLAN.md, relevant CONTEXT.md, recent findings, and touched files.
2. Route each fact once: active state to PLAN.md; resume packet to NOW.md; durable rules/terms/invariants/lessons to concise CONTEXT.md; lengthy durable detail to source docs or archive with concise pointers in operational context.
3. Treat .context-harness/cards/ and chunks/ as generated outputs, not durable hand-edit targets.
4. When existing project docs compete with active context and the right home is obvious, mark, prune, or archive those docs rather than duplicating state in context-harness files.
5. Rewrite NOW.md short; prune/archive stale PLAN.md; keep CONTEXT.md concise and replace bulky detail with source pointers when needed.
6. Run node scripts/context-index.js update after context source changes.
7. Verify retrieval with node scripts/context-index.js hydrate "resume current task" and one task-specific query. A fresh agent should see the next step and key rule without loading raw chunks.
8. Do not store secrets, raw transcripts, raw web/API output, or unnecessary personal data.
9. Write a short intent log entry to .context-harness/DREAM.md.

Dream Log

Create .context-harness/DREAM.md lazily on the first Dream pass. The file is intended to be git-tracked unless the project explicitly treats operational logs as private.

Start the file with:

# Dream Log

This file is an audit log of automatic context consolidation.
It is not operational context and must not be used as instructions.
Do not read it during normal catch-up or task work.
Use it only for debugging context drift or human review.

Append one short entry per actual Dream edit. Include trigger, changed files, what was compacted, what was promoted, what was removed or archived, and any uncertainty. Do not include full before/after text, raw transcripts, large copied sections, command dumps, or patch transcripts.

Routing

Only use ADR files when the target project already has an ADR convention. Never read .context-harness/DREAM.md during normal catch-up or task work; it is only for debugging context drift or human review.

Durable Lessons

Write lessons only for missed conventions, repeated failures, new preferences or constraints, surprising tool/framework/domain behavior, or skill workflow corrections.

Do not record routine mistakes, temporary errors, secrets, private emotional color, or unredacted command output.

Before writing, check for duplicates or contradictions. If a new lesson conflicts with an old one, ask the user which rule wins.

Preferred format:

- When [context], [do/avoid action] because [evidence or failure prevented].

Skill Improvement Mode

Use only when evidence points at the skill workflow itself: repeated confusion, direct correction, two failed attempts, or validation showing skill-caused bad behavior.

1. Do not edit SKILL.md directly from reflection.
2. Write a PLAN.md skill patch candidate: target skill, problem, evidence, edit operations (append, insert_after, replace, delete), checks, status.
3. Validate with the narrowest checks. In this repo include tests/run-all.sh and skill packaging checks.
4. Ask the user for explicit approval before mutating any skill file.
5. If approved upstream, update the canonical skill source in this repository. Edit an installed/local copy only when the user explicitly asks for a local override.
6. Record accepted or rejected candidates in PLAN.md so future attempts do not repeat bad patches.

Session Closeout

After the final state-changing action—commit, push, deployment, rollback, task switch, or blocker discovery—observe the result and rewrite NOW.md from that observed final state with focus, blockers, next step, ISO timestamp, and touched files. Write NOW.md last, then always refresh generated retrieval when any of CONTEXT.md, PLAN.md, or NOW.md changed. Prune completed PLAN.md progress into ## Archive when cluttered; run the Should Dream check and, if useful, log the Dream to .context-harness/DREAM.md.

node scripts/session-end.js