Save Context

Save Context

Save current session state to CONTEXT-{stream}-llm.md with LLM-optimized format.

Target: 2500-3000 tokens MAX | Speed: 3-5 seconds (Not a throwaway summary: the working memory of a multi-session stream, holding accumulating per-role lines.)

⚠️ AskUserQuestion Guard

CRITICAL: After EVERY AskUserQuestion call, check if answers are empty/blank. Known Claude Code bug: outside Plan Mode, AskUserQuestion silently returns empty answers without showing UI.

If answers are empty: DO NOT proceed with assumptions. Instead:

1. Output: "⚠️ Questions didn't display (known Claude Code bug outside Plan Mode)."
2. Present the options as a numbered text list and ask user to reply with their choice number.
3. WAIT for user reply before continuing.

⚠️ VERBATIM RULE

CRITICAL: Never rewrite the text after the colon of a carried-forward line. Never touch a line whose role is not yours — copy it byte-for-byte. A Haiku pass once reworded a user's exact phrasing while explicitly asked to quote it verbatim — hence Sonnet here, not Haiku.

Performance Rules

1. Use rtk for ALL shell commands
2. Parallel tool calls — ALL independent calls in one message
3. Minimize round-trips — gather all data phase 1, reason phase 2, write phase 3

Workflow

Phase 1: Gather Data (parallel)

Bash: rtk ls openspec/changes/ + rtk ls -t CONTEXT-*llm.md
Read: existing CONTEXT-{stream}-llm.md (if present) — needed to carry forward owned lines and to read the `saved:` field for the Phase 3 concurrency check

Stream resolution: First word of $ARGUMENTS = stream name (^[a-zA-Z0-9_-]{1,50}$), rest = description. Empty → reuse prior /load-context stream or AskUserQuestion.

Role resolution: --as <role> flag in $ARGUMENTS > role recorded by the last /load-context for this stream > none (solo session — untagged line format, unchanged behavior).

Phase 1b: Detect Thinking Artifacts (parallel with Phase 1)

If $PRAXIS_DIR is set:

Bash: ls -t "$PRAXIS_DIR/thinking"/*/{project}/ 2>/dev/null | head -10

Where {project} = current project folder name. Collect recent artifact paths written during this session (match conversation timestamps/topics).

If $PRAXIS_DIR is unset or empty: skip silently — no error, no warning.

Phase 2: Collect Trailers & Synthesize

Collect-then-fill. Trailers (<!-- ckpt ... -->) are author-distilled at the moment of decision — route them mechanically, and only synthesize what they don't cover.

1. Collect — gather every <!-- ckpt block already present in this session's own conversation. No transcript parsing, no scripts, no greps of ~/.claude — the trailers are already in context.
2. Route — map each clause to its target section via the routing table in reference.md. Mechanical, no rewording.
3. My lines = my old lines carried forward BY COPY (verbatim, from the Phase 1 Read) + my new trailers + my ## Hot Files rows re-derived fresh (that list is regenerated every save — but MY rows only; Hot Files is an owned section like the rest). Pruning:
   • A reversed decision is NOT deleted — move it to ## Dead Ends WITH its rationale intact.
   • An answered open: is removed.
   • Criterion is VALIDITY (still true?), never line type — never drop a reasoning: for not being a decision.
4. Other roles' lines — copy byte-for-byte. Never edit, reorder, reword, or reflow a line whose role isn't mine, their ## Hot Files rows included: "fresh each save" governs my rows, it is not license to overwrite theirs.
5. Synthesize — ONLY what trailers don't cover: focus, goal, Next, status, Drop, progression.
   • Next — union of routed open: lines + inferred tasks. IMPORTANT: heading must stay Next (Claude Code compaction grep-matches next/todo/pending/remaining for survival priority).
   • Thinking Artifacts (if Phase 1b found any) — paths only, no content.
   • Predecessor — path of the CONTEXT this session resumed (/load-context) or overwrites, so streams chain into a walkable history. Else none.
6. Fallback — zero trailers this session → pre-rework behaviour: infer Session (decisions/thinking/unexpected), Learnings, Dead Ends from the last 15-20 messages.

Phase 3: Write & Report

Before writing, apply three pre-write passes:

• Concurrency check: re-read the saved: field from the on-disk CONTEXT file. If it changed since Phase 1's read, another session wrote in between — re-read the file, re-merge (Phase 2 steps 3-4), and retry. Optimistic check; no lock file.
• Reference-don't-copy (named contract): never inline content that already lives in a spec/plan/diff/OpenSpec/thinking-artifact — reference it by path. The CONTEXT file points at sources; it does not duplicate them.
• Scrub secrets: strip anything key/token/credential/PII-shaped from the synthesized text (API keys, bearer tokens, KEY=/PASSWORD= values, private-key blocks) → replace with [REDACTED]. The resume file must never persist a secret.

Write CONTEXT file using template.

Stream naming: "default" → CONTEXT-llm.md, "{name}" → CONTEXT-{name}-llm.md

Report — growth line so the human can watch accumulation: 📈 {total} lignes · ~{tok} tok · [{role} {n}] [{role} {n}] (+{delta} depuis {last_save})

Phase 3b: Auto-archive to done/

If status is done or parked → move file to done/ subfolder:

Bash: mkdir -p done && mv CONTEXT-{stream}-llm.md done/

Report: "📦 Archived to done/ (status: {status})"

See reference.md for CONTEXT file template, quality self-check, and done/ archival rules.