Llmdoc Update

llmdoc-update

This skill is the Codex-native equivalent of /llmdoc:update.

The update target is tracked llmdoc/, not .llmdoc-tmp/. Stable docs should stay consistent with the current repository and remain smaller than the source code or provide architectural explanation that source search does not.

Change detection is commit-based: it is anchored on a durable watermark (llmdoc/state/sync.md) recording the last source commit already reflected in llmdoc/, not on the working tree or the date. Uncommitted changes are an additional input; they never move the watermark.

Use it when:

• a task changed project knowledge, architecture understanding, or workflow guidance
• a useful mistake or missing-doc lesson should be preserved
• you want a command-like Codex entrypoint for updating llmdoc

Read the batch flags from $ARGUMENTS / the user message: [summary] [--range A..B ...] [--commits SHA,SHA ...] [--since REF] [--from SHA] [--working-tree-only] [--include-default]

Invariants, design rationale, and verified git semantics: llmdoc/architecture/update-orchestration.md.

Before editing stable docs:

• read llmdoc/index.md, llmdoc/startup.md and the MUST docs it lists
• proactively read relevant llmdoc/guides/ and llmdoc/memory/reflections/
• read skills/llmdoc/references/lessons-learned.md for the active-memory archive threshold
• check relevant .llmdoc-tmp/investigations/ reports only as local temporary evidence (validate their git revision, resolved range, scope, gaps)
• align with the user before non-trivial edits

Sync State

llmdoc/state/sync.md is a tracked, machine-managed markdown file holding one load-bearing field, watermark-commit (the full commit SHA, whatever hash the repo uses — SHA-1 or SHA-256). It is not project knowledge: never index it, never add it to startup.md/must/, never count it as active memory. Read the watermark with a single anchored line:

W=$(awk -F': ' '/^- watermark-commit:/{print $2}' llmdoc/state/sync.md 2>/dev/null | tr -d '[:space:]')

recorder is the only writer of sync.md, and only advances it as the terminal step of a successful update.

Change Detection (commit range)

Run all git plumbing without set -e (capture each exit code) so a 128 degrades instead of aborting. The resolution ladder order is an invariant — never reorder:

# 1. Capability probe (gates the whole ladder — steps 2-5 assume a git work tree)
[ "$(git rev-parse --is-inside-work-tree 2>/dev/null)" = true ] || {
  echo "non-git: watermark inactive"   # STOP HERE: use legacy working-tree + staged + summary detection;
  return 0 2>/dev/null || exit 0        # do NOT run steps 2-5, do NOT read or advance the watermark
}
SHALLOW=$(git rev-parse --is-shallow-repository)
H=$(git rev-parse HEAD)                                  # capture once

# 2. Read watermark
W=$(awk -F': ' '/^- watermark-commit:/{print $2}' llmdoc/state/sync.md 2>/dev/null | tr -d '[:space:]')

# 3-4. Resolve RANGE_BASE with EXPLICIT branching. Existence gates reachability: never run merge-base
#      on an empty/missing $W — it would fatal-128, not skip. Each branch is terminal.
if [ -z "$W" ]; then
  echo "watermark empty/missing → first-run: base=HEAD (no backfill unless --from/--since), force ≥ analysis, seed at HEAD on success"
  RANGE_BASE=$H
elif ! git rev-parse --verify --quiet "$W^{commit}" >/dev/null; then
  echo "watermark object missing (GC'd / below shallow boundary) → first-run baseline=HEAD, warn, do NOT advance"
  # if $SHALLOW = true, suggest: git fetch --unshallow. (rev-parse --verify --quiet exits 1 cleanly; cat-file -e "$W^{commit}" would fatal 128 on the peel.)
  RANGE_BASE=$H
elif git merge-base --is-ancestor "$W" "$H"; then
  RANGE_BASE=$W                                          # watermark valid
elif git merge-base --is-ancestor "$H" "$W"; then
  echo "HEAD BEHIND watermark → REFUSE (reverse W..HEAD would list future files as deletions and strip docs); working-tree input only, do NOT advance"
  RANGE_BASE=$H
else
  RANGE_BASE=$(git merge-base "$W" "$H")                 # true divergence (rebase/squash-merge); force ≥ analysis
fi

# 5. Net change set (rename-aware, NUL-safe). In -z mode a rename/copy record is STATUS\0OLD\0NEW\0
#    (3 fields; STATUS is R<score>/C<score>); a normal record is STATUS\0PATH\0.
git diff --name-status -M -C -z "$RANGE_BASE" "$H"

Batch flags (if any of --range/--commits/--since/--from is given, the default RANGE_BASE..HEAD is NOT added unless --include-default). Each batch resolves to a set of changed paths: parse the --name-status -z records — drop the status column; in -z mode a record is STATUS\0PATH\0, but a rename/copy record (status R<score>/C<score>) is STATUS\0OLD\0NEW\0, so read TWO path tokens when the status starts with R or C (take both) and one otherwise, or the parser desyncs. Union all batch path-sets with the working-tree set, deduplicated (do not sort -u raw NUL/status records):

• --range A..B (repeatable): git diff --name-status -M -C -z A B.
• --commits SHA,…: per commit git diff --name-status -M -C -z <sha>^ <sha>; reject merge commits (parent count > 1 → ask the user to pass A..B); root commit diffs against the empty tree git hash-object -t tree /dev/null (4b825dc642cb6eb9a060e54bf8d69288fbee4904).
• --since REF|date: for a ref, <ref>..HEAD; for a date, resolve a boundary commit B=$(git rev-list -1 --before=<date> HEAD) and diff B..HEAD with the same git diff --name-status -M -C -z as every other branch. If B is empty (no commit predates <date>, so the whole history is in range), diff against the empty tree git hash-object -t tree /dev/null instead of B..HEAD (which would be ..HEAD and fatal). Do NOT git log --name-only, which yields a statusless, blank-line-polluted shape that will not union cleanly.
• --from SHA: one-off start override for this run only; does not touch sync.md.
• --working-tree-only: skip the committed range entirely; document only uncommitted changes; never advance the watermark.

Always fold in, as an ADDITIONAL input tagged worktree (never advances the watermark): git diff --name-only HEAD (every tracked path whose on-disk content differs from HEAD — staged and unstaged) ∪ git ls-files --others --exclude-standard (untracked). git diff --name-only HEAD already covers staged changes, so a separate --cached diff is redundant.

Empty range (already synced) — applies ONLY when the watermark was VALID (the RANGE_BASE=$W branch) and W == HEAD: if there are also no working-tree changes and no batch flags, report already up to date through <Hshort> and exit without touching docs, the index, or the watermark. The degraded branches that force RANGE_BASE=HEAD (empty/missing watermark → first-run seed; HEAD-behind → refuse) also produce an empty range but are NOT "already synced" — each is handled by its own branch outcome above, not by this terminal.

Doc-commit loop-breaker: if git diff --name-only "$RANGE_BASE".."$H" -- ':(exclude)llmdoc/**' is empty (range non-empty but doc-only), fast-forward the watermark to HEAD with zero doc work. Always exclude the sync file and .llmdoc-tmp/ from every diff.

Content selection: recorder reads committed-batch file content at the BATCH TIP (git show <tip>:<path>), never from HEAD/disk, except for the uncommitted working-tree set (read from disk, which wins for any currently-dirty path).

Mode Selection

The trigger is range size × authorship × risk (not context freshness):

• fast: small range (≤ ~3 commits), self-authored, impacted docs nameable. --working-tree-only defaults here. Use the task summary, diff, targeted checks, and any still-valid scratch reports.
• analysis: ~4–15 commits, OR any non-self-authored commit, OR multiple clusters, OR a recovered/derived/first-run baseline. One focused evidence pass persisted under .llmdoc-tmp/investigations/.
• full: > ~15 commits, multi-batch backfill, history-rewrite recovery, disputed facts, or reflection-heavy failures — separate investigation, reflection, and recording roles.

Hard floors (force ≥ analysis): a merge-base-recovered baseline, a derived/first-run baseline, or any non-self-authored commit. Backfill blast-radius cap: a first-run/--since range beyond ~20 commits or ~50 files forces full and explicit user confirmation.

Self-authored test: a commit is self-authored when its author email (git log -1 --format=%ae <sha>) equals the current git config user.email. If the identity cannot be determined (e.g., a subagent with no git identity, or user.email unset), treat commits as non-self-authored — the conservative side that forces ≥ analysis.

Workflow

1. Rebuild task context (index, startup, MUST docs, relevant guides + reflections; note any $ARGUMENTS summary).
2. Resolve sync state and compute the change set (run the ladder; parse batch flags; union with the working-tree set; apply the loop-breaker; handle degraded cases — non-git, shallow, first-run, orphaned, diverged, HEAD-behind — without fabricating a watermark advance).
3. Select the mode from range size × authorship × risk; honor the hard floors and backfill cap.
4. Investigate only as needed, seeded with the resolved net-diff path list; scratch reports record the resolved RANGE_BASE..H range.
5. Reflect only when there is a workflow failure, repeated mistake, missing signal, or durable process lesson. Do not force a reflection for routine fast updates.
6. Update stable llmdoc docs against the batch-tip state: update only impacted docs, correct stale claims, split aggressively, reconcile llmdoc/memory/doc-gaps.md.
7. Run the active-memory archive check (count files under llmdoc/memory/ excluding lessons-learned.md, doc-gaps.md, archive/; llmdoc/state/ is not counted). If > 5, follow skills/llmdoc/references/lessons-learned.md.
8. Synchronize llmdoc/index.md. Do not index .llmdoc-tmp/, and do not index llmdoc/state/sync.md as knowledge.
9. Advance the watermark (recorder-owned terminal step). Safe-to-advance gate — ALL must hold: the update completed successfully and consumed a committed range; HEAD is attached (git symbolic-ref -q HEAD succeeds); and no git operation is in progress — test by whether the resolved path EXISTS on disk (git rev-parse --git-path always prints a path and exits 0 regardless of existence, so check with [ -f ]/[ -d ]), none of [ -f "$(git rev-parse --git-path MERGE_HEAD)" ], CHERRY_PICK_HEAD, REVERT_HEAD, [ -d "$(git rev-parse --git-path rebase-merge)" ], rebase-apply may exist. If the gate holds, advance watermark-commit to the captured H (or the highest unbroken-prefix tip). Rewrite ONLY these fields, keeping the exact - watermark-commit: line prefix (the reader anchors on it — do not reformat): watermark-commit (new full commit SHA), watermark-subject (git log -1 --format=%s <new-sha>), updated-at (ISO-8601 UTC, date -u +%Y-%m-%dT%H:%M:%SZ), updated-by (/llmdoc:update). NEVER advance on a --working-tree-only run, a failed/partial run, a HEAD-behind-watermark run, or when the safe-to-advance gate fails.
10. Report the mode used, resolved range(s)/batches and commit count, old → new watermark (or why it did not move), scratch/reflection paths, the archive action, and the stable docs that changed.

At the end of a non-trivial task, proactively consider whether the user should be prompted to run this workflow.