Skills pin via a submodule at .claude/appland-skills (the pattern used
in nova-2), with the project's .claude/skills entries symlinked into
it — replacing machine-local symlinks into a sibling checkout that only
resolved on one workstation. The submodule SHA records which skills
revision the repo's gold-traces baselines were maintained with.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…rride

The action shipped as v1 (usage reporting, cached toolchain install,
live agent progress, per-matrix sticky comments via comment-tag — which
this two-package matrix needs). The skills feature branch this workflow
pinned is merged to skills main, the action's default.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Retroactively apply the new 'appmap sanitize' tokenization to all 20
committed baselines: every captured value string becomes a per-file
equality-preserving token, so the committed traces are structurally
incapable of carrying a secret. 17% smaller as a side effect
(1,027,715 -> 852,294 bytes).

Verified digest-safe against the gold-traces engine's own pipeline:
the bless digest (sha256 over the exported sequence diagram's
rootActions subtreeDigests) is identical before and after for all 20
baselines, so no re-bless and no false drift. Sanitization verified
idempotent on every file.

Also removes a stale manifest comment referencing the pre-engine
gold_traces/trim.mts script.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Three gold traces for one subsystem multiplies review and bless cost
without strengthening the baseline (the skill's own guidance: keep
one). A new composite unit test walks every sanitization branch in a
single AppMap — shared-namespace masking, headers/paths, SQL both
parseable and not — and its recording replaces the three baselines.
The recording is sanitized, matching what the engine will do once it
switches from trim.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

A skipped file keeps its original, possibly secret-bearing contents, so
a partial batch must not look like success — a CI gate trusting the
exit code could otherwise commit an unsanitized AppMap. All files are
still processed before failing (stopping at the first bad file would
leave the rest untouched); best-effort callers can append '|| true'.
Handler-level tests cover both outcomes; the gold trace is unchanged
(one composite trace per subsystem).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…push

A review costs a few dollars of agent time; pushes no longer trigger
one. Add the appmap-review label to review a PR (remove and re-add to
review again). The per-package change check and the fork guard stay;
workflow_dispatch stays as the manual path and the way to refresh the
baseline on main.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

workflow_dispatch gains head and base inputs (defaults: the picked
branch, main). The checkout, the per-package change check, and the
review's base/head revisions all honor them, so any two refs can be
re-reviewed against each other — and a dispatch against an explicit
base now also skips packages the head didn't change.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Each command now records what it did in the file's metadata:

  metadata.sanitized: { version, allow_values }
  metadata.trimmed:   { version, max_length }

Behavior on a marked file: skip when the same CLI version already ran
with the same options; re-apply otherwise. Re-applying only tightens —
masking and truncation are one-way — so the marker records what still
holds: the allowlist intersection of all runs, and the smallest cap.
Both commands warn when options loosen (re-record to actually loosen).

The marker also replaces in-band token detection: in a marked file the
tokens are ours and pass through; in a fresh file a string that merely
looks like a token is application data and is masked — closing the
last gap where a lookalike value could slip through or alias a token.

All 22 committed baselines are marked (they were sanitized by this
code); no values changed.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The review agent blessed it via the skills-main engine, which still
trims; sanitize it and give it the provenance marker like the rest.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The handler now runs buildAppMap().source().normalize().build() before
masking — the same transform every AppMap viewer applies on load. What
that adds: credentials are stripped from metadata.git.repository (a
capture surface the masking walk didn't cover), eventUpdates are merged
into their events, event ids are re-indexed, and unreturned calls are
balanced. Masking then runs over the canonical form.

Note: normalize() does NOT parameterize SQL or group routes — sanitize's
own fail-closed SQL handling remains the only thing scrubbing SQL.

All 24 committed baselines re-processed through the full pipeline;
bless digests verified identical for every one.

Co-Authored-By: Claude Fable 5 <noreply@appland.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

metadata carries captured data too: git identity (user_name /
user_email), credentials embedded in the repository URL, and exception
/ test-failure messages. The command pipeline's normalize step also
strips repository credentials, but the exported sanitizeAppMap must
not depend on the command wrapping it — the security guarantee lives
in the walk, alone. An unparseable repository URL is masked whole
rather than risking an embedded credential.

Branch, commit, and code locations stay verbatim: repo state, not
captured data.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>