Backfill

Reconstruct an OKF bundle from history

This skill replays a repository's decision-making history (git commits and Claude session transcripts) to rebuild its OKF knowledge bundle as if the okf skill's Stop hook had been active from the start.

The extraction layer is deterministic (same repo → byte-identical events); the replay layer is an LLM loop, replayable and auditable but not byte-identical (timestamps, summaries change per run). See the event schema (§1) and skip rules (§3).

1. Event schema

Git commits and session turns become events (JSONL, one per line, sorted by timestamp then source):

{"id":"git:<sha>","source":"git","ts":"<ISO8601 Z>","sha":"...","author":"...","subject":"...","body":"...","files":[{"path":"...","add":N,"del":N}]}
{"id":"session:<file>:<lineno>","source":"session","ts":"<ISO8601 Z>","user":"...","outcome":"...","title":"...","branch":"...","skip":"..."}

• Git events come from git log --first-parent --numstat.
• Session events pair each user message with the last assistant text block of that turn (the wrap-up), extracted from ~/.claude/projects/<repo-slug>/*.jsonl and worktree subdirs.
• Timestamps are normalized to UTC ISO 8601 strings ending in Z.
• Skip field (optional, added by extraction): marks low-signal events — see §3, never overridden by replay.

2. Protocol: Preflight → Extract → Bootstrap → Replay (Map+Reduce) → Finalize

Preflight: check if bundle can be rebuilt

if [ -d <repo>/.okf ] && ! [ -f <repo>/.okf/.backfill-state.json ]; then
  echo "ERROR: .okf/ exists but is incomplete. Delete it to rebuild from scratch, or pass --resume to continue from the last checkpoint."
  exit 1
fi

If .backfill-state.json exists, resume from the cursor; otherwise, fresh bootstrap.

Extract: generate event stream

uv run "${CLAUDE_SKILL_DIR}/scripts/okf_backfill_events.py" <repo-dir> \
  --out events.jsonl \
  [--no-sessions] [--sessions-dir ~/.claude/projects] \
  [--max-text 2000] [--skip-globs "vendor/**"]

Write events.jsonl to the scratchpad (never committed). Report:

• Total events extracted, per-source counts, and per-rule skip counts.

Bootstrap: initialize bundle (fresh only)

Create .okf/ with:

mkdir -p <repo>/.okf
echo 'okf_version: "0.2"' > <repo>/.okf/index.md
echo '# Update Log' > <repo>/.okf/log.md

Cursor (.okf/.backfill-state.json):

{"last_id": null, "done": 0}

Replay: two-phase map/reduce protocol

The replay is now structured in two phases to enforce semantic depth and anti-degeneration rules:

Phase 1: Map (parallel analysis, per-event)

Launch okf:event-analyzer agents for the live events (those without skip field), in waves of 4 to 10 (see the cost note in §5), via Claude Code's Workflow agentType parameter. Each analyzer receives its event ids, the events.jsonl path, the repo path, the output directory and the emitter path, and:

• Fetches its own event JSON with jq (the orchestrator dispatches ids, never content)
• Reads git evidence only through the capped diff emitter (§6); session events need no other call
• Writes analyses/<event-id>.md (event id with : sanitized to -) to scratchpad, with a truncated flag copied from the emitter's last line
• Replies with one line of counts per event; the analysis never travels back in the reply
• Never touches .okf/
• Is resumable: skip events already in analyses/

Dispatch rule (deterministic, from events.jsonl, no content read): a live event is small when it is a session turn or a commit whose numstat totals at most 60 changed lines. Small events are grouped chronologically, eight per analyzer call; large commits go one per call. Measured on this repository (benchmark/map-tier/RESULTS.md): 57 calls instead of 147, a third off the map phase at the same tier, no loss on any per-event metric and no cross-event contamination.

jq -c 'select(.skip==null) | {id, small: (.source=="session" or (([.files[]?|.add+.del]|add)//0) <= 60)}' events.jsonl

Host-agnostic fallback (if no Workflow support): spawn generic subagents with the same system prompt as agents/event-analyzer.md, one per event, collecting analyses to scratchpad.

Phase 2: Reduce (sequential folding, one agent)

Run a single okf:bundle-weaver agent that:

• Reads all analyses in chronological order
• Folds them into .okf/, updating or creating concepts
• Enforces anti-degeneration rules (§4)
• Adds a dated log.md bullet for each lifecycle event only (a concept created, deprecated or superseded); the "why" of a routine update goes into the concept body
• Manages cursor (.okf/.backfill-state.json) for resume capability
• Is the only actor that writes .okf/
• Replies with one line of counts (folded, created, updated, bullets, conflicts, superseded, truncated inputs); the bundle never travels back in the reply

Resume behavior: both phases support resumption. Phase 1 skips already-analyzed event ids; Phase 2 restarts from last_id in the cursor.

Domain priming (advisory)

Before starting the map phase, read the repository's README and directory structure to sketch a candidate taxonomy of concepts (skills, integrations, decisions, etc.). This priming is advisory only — it helps the analyzer emit better candidate names. The guarantee that no events are lost comes from the deterministic coverage check in Finalize (§5), not from priming; analyses without candidate names still flow to the weaver.

Context hygiene (orchestrator)

The orchestrator handles ids and counts, never content. Events, analyses and diffs are read by the agents; the orchestrator reads:

wc -l events.jsonl                                   # total events
jq -r 'select(.skip==null) | .id' events.jsonl       # live ids to dispatch
ls analyses | wc -l                                  # analyses written

Never cat events.jsonl or open an analysis from the orchestrator: whatever enters its context is billed at the frontier rate on every following turn, and the routing exists to prevent that.

Finalize: validate and clean up

1. Write index.md per directory by hand (do NOT run okf_init.py — it would scaffold placeholder files): one # Section per directory, one bullet per concept, * [Title](file.md) - description taken from each concept's frontmatter (SPEC §8). The root index.md keeps its okf_version: "0.2" frontmatter and links every subdirectory.

2. Add final log entry (today's date, appended to existing dated section if present):

   ## 2026-09-01
   
   - **Backfill**: reconstructed from N events by okf-backfill/0.9.5

3. Delete cursor:

   rm -f .okf/.backfill-state.json

4. Run self-checks for anti-degeneration rules:

   # No concept filenames derived from change type
   ! find .okf -name "*.md" | grep -E "merge-pull-request|(^|/)(feat|fix|chore|docs)[:-]"
   
   # No non-kebab-case characters in filenames
   ! find .okf -name "*.md" | grep -E "[^a-z0-9/.-]"
   
   # No identical consecutive log bullets
   awk 'p==$0 && /^- / {exit 1} {p=$0}' .okf/log.md
   
   # Every concept declares its lifecycle (absent status reads as `stable`, SPEC §5.4)
   [ -z "$(grep -rL '^status:' --include='*.md' .okf | grep -v '/index.md$\|/log.md$')" ]

   These are lexical guards; they do not catch a bundle that asserts a superseded state. That check is semantic and belongs to the weaver, at fold time, when it has both the analysis and the earlier concepts on disk (§4 rule 5).

5. Run coverage check to guarantee every event is mapped:

   uv run "${CLAUDE_SKILL_DIR}/scripts/okf_backfill_events.py" \
     --check-coverage events.jsonl .okf

   Exit code 0 means all live events (non-skipped) appear in the bundle's sources or log. Exit code 1 with a list of unmapped ids means the weaver skipped some events; fix and re-run.

6. Validate bundle schema:

   uv run "${CLAUDE_SKILL_DIR}/../validate/scripts/okf_validate.py" .okf --strict

   Fix every error before finishing.

7. Report the resulting bundle:

   • Concept count per directory
   • Log entry samples (first and last)
   • Coverage check result (all events mapped)
   • Validation result (pass/fail, warnings)
   • Superseded resolutions (the weaver's superseded count) and deprecated concepts: grep -rl '^status: deprecated' --include='*.md' .okf | wc -l
   • Agents spawned per phase (analyzers, weaver invocations) and, when the host reports it, tokens per phase
   • Truncated analyses: grep -l '^truncated: true' analyses/*.md | wc -l, next to the deterministic estimate of git events whose full diff exceeds the default cap: jq -r 'select(.skip==null and .source=="git") | [.files[]|.add+.del] | add' events.jsonl | awk '$1>300' | wc -l

3. Skip rules

Events matching a rule below are marked skip: <rule-id> and skipped during replay (cursor advances, no log entry):

Extend the first rule with --skip-globs "extra/**" for repo-specific patterns.