fix(guard): resolve cognition paths against the session dir, not the anchor

[ProtocolWarden]

The defect

ContextGuard joined capsule_path / checkpoint_path / handoff_path onto CL_ANCHOR, while .context/config.yaml documents them as relative to .context/sessions/<CL_SESSION_ID>/. CL_SESSION_ID appeared nowhere in either hook — 11 join sites across pre_tool_use.sh and stop.sh, every one resolving to a directory that does not exist.

Consequences, all silent:

• lease expiry never fired — the handoff was never located
• worker forbidden_paths never loaded
• require_capsule: true was a landmine, not a feature — it would have blocked every tool call with "No active capsule found" rather than enforcing anything

The guard verified CL_ANCHOR was set and then, in practice, did nothing.

Proof, not inspection

A fixture anchor carrying an expired lease:

old hook: exit 0   (silently passed)
new hook: exit 2   {"decision":"block","reason":"...Lease expired at 2020-01-01T00:00:00Z..."}

Backward compatibility

Both hooks resolve against a new COGNITION_ROOT. When CL_SESSION_ID is unset it falls back to the anchor's .context/ and warns, rather than blocking.

Path defaults drop their .context/ prefix to match the config convention. A consumer whose config still pins the old .context/active/ values would otherwise have resolved to <anchor>/.context/.context/active/, so configured values now have a leading .context/ stripped. Verified: a legacy layout + legacy config with no session id blocks identically on both the old and new hook.

Also fixed

block() hand-built its JSON, so a reason containing a Windows path emitted an unescaped backslash (\U) and the payload failed to parse — a blocked call reached the operator with no reason attached. Now serialized with json.dumps.

Testing

• tests/test_hook.py + tests/test_pre_tool_use.py: 26/26 pass
• Full suite: 17 failures, unchanged from main on this machine (Windows environment: RepoGraph registry, fcntl locking). Not introduced here.

Reviewer note

The CL_SESSION_ID is not set warning fires per tool call, matching the existing warn() call sites in this file, which behave the same way under persistent conditions.

🤖 Generated with Claude Code

[ProtocolWarden]

Review pass — two defects found in this change and fixed before merge.

1. Backward-compat break (would have shipped). Dropping the .context/ prefix from the path defaults is correct for the new base, but a consumer whose config.yaml still explicitly pins the old .context/active/ value would have resolved to <anchor>/.context/.context/active/ — silently re-creating the exact bug this PR fixes, for the users least likely to have migrated. Fixed by stripping a leading .context/ from configured values, and verified against a legacy-layout fixture: identical block behaviour on old and new hooks.

2. stop.sh would have crashed under set -u. The bulk rewrite introduced COGNITION_ROOT references into stop.sh before adding its definition — an unbound variable on every session end. Caught by syntax + functional check, definition added.

Checked and deliberately left alone:

• Template: .context/templates/... in stop.sh stays anchor-relative — templates are manifest-wide, not per-session. Correct as-is.
• The SESSION_MARKER hash keys on REPO_ROOT only, so two session ids under one anchor share a marker. Pre-existing, out of scope, not made worse here.
• The CL_SESSION_ID warning fires per tool call. Matches the existing warn() call sites, which behave identically under persistent conditions.

Hook tests 26/26. Full-suite failure count unchanged from main.

[ProtocolWarden]

CI triage — red, and provably not from this change

Lint (ruff) fails with 205 errors. This diff cannot be the cause:

$ git diff --name-only main...fix/session-scoped-cognition-paths
CHANGELOG.md
adapters/claude/hooks/pre_tool_use.sh
adapters/claude/hooks/stop.sh

Zero Python files. Ruff only checks .py. Confirmed locally too — ruff check src reports an identical 90 errors on main and on this branch, and ruff check . is identical on both.

The flagged file in the log is a test module's import block, untouched here.

Probable root cause, worth its own issue

.github/workflows/ci.yml:33 installs ruff unpinned:

run: pip install "ruff>=0.5"

main last went green on 2026-07-17. Nothing since then has run CI on main itself. An unpinned linter means a new ruff release turns previously-green code red with no commit at all — which matches 205 errors appearing on a repo that was clean in July. Pinning ruff (or --exit-zero on legacy debt with a ratchet) would stop every future PR inheriting an unrelated red.

Test (pytest), License headers, and audit all pass on this branch.

Merging on that basis: the failure is pre-existing and independent of this diff.