TASK-20 / TASK-5238 catalog handoff

Session: 4f00; actual thread: 01a080f6-e8f2-7032-a4e4-2204c189d088. Worktree: .openmates-agent-worktrees/agent-4f00, distinct from coordinator agent-4aa6 and other workers. Task thread connection is pending; API attempts returned HTTP 429. Per coordinator, no further task reads/connect/posts until serialized recovery.

Implemented removal patch

• Removed both urban sports examples, dermatology, PDF encryption, product-launch mindmap, old career, OpenMates get-docs, private-workspace demo video: eight fixtures, translations, registry entries and interest references.
• Removed those landing references and missing learning ID. Local-AI compression and Svelte remain in the catalog outside landing.
• Existing workshops remains mapped. Other slide arrays await approved candidates. No replacements registered and no candidate approval inferred.
• backend/apps/openmates/app.yml: sparse default_enabled: false for openmates.get-docs. Separate code.get_docs remains enabled.
• Removed rejected-fixture-only mindmap test and video fixture cases. Updated ranking fixture to preserved local-AI; settings shell test to preserved Svelte; focus leak test to existing Framework reputation. Shared focus/video implementation stays with assigned owners.

Verification and remaining gates

• SDK/CLI parity in this worktree: 89 commands and 4 workflow templates passed.
• git diff --check passed.
• Earlier focused checks passed 17 tests and generated settings metadata excluded openmates.get-docs, but ran from canonical checkout containing this exact patch before routing recovery. Worktree rerun pending dependency bootstrap.
• Earlier structural audit identified one content gap: mindmaps.mindmap has no example after requested rejection. Keep visible until accepted replacement; do not weaken audit.
• Specification source-path check passed; changed-test check reported stale generated assertion index. Recheck against current coordinated Specification artifacts before integration.
• Queued CI request fae4df5ff544e058c39969522cfa4246bb4050d29e5e117b3cdaf635d8f5d95c, candidate dad7eefcc84ff2a68060daac5130d569275af651, is INVALID as patch evidence: created before routing recovery, with no implementation diff. Do not use its result for acceptance.
• No dev deployment, reviewed browser/video proof, live runtime restart or task completion claimed. Runtime changes require explicit target and coordinator lease.

Replacement handoff required

Coordinator/example owners must supply candidate source/i18n paths and passing review records (real CLI, phone/laptop, required speech admission), then catalog owner can integrate. None supplied yet. Mindmap category currently lacks an admissible example. Keep assignment open.

Routing recovery

Final directory check found default exec calls had used canonical root despite session binding. Stopped writes; inspected exact root-dirty metadata; imported all 27 reviewed task paths with sessions.py worktree import-root; checked identical bytes; reversed only the exact task patch in canonical root. Recovery backup: /tmp/task20-root-routing-recovery.patch. Canonical root retained only an unrelated untracked config.json, not read or staged. Every subsequent command sets explicit workdir to agent-4f00.

Agent process retrospective: repeated Tasks API attempts before coordinator backoff and an unsupported skill-prescribed --dry-run wasted calls. More seriously, relying on binding without explicit exec workdir required patch recovery and invalidated CI evidence. Existing session instructions already require isolation; recommend a deterministic pre-exec hook/audit that rejects mutating shell calls whose actual cwd differs from the bound worktree. Inspection confirms scripts/codex_hook_context.py route() handles Bash/bash and apply_patch, but returns unchanged for other tool names. Smallest hook fix: support exec_command cmd/workdir payloads, including orchestrated calls, with a regression test that a bound chat cannot mutate canonical root. No product-scope workflow tooling edits made.

Verified worktree rerun

Offline frozen-lockfile dependency installation completed in agent-4f00. Focused Vitest rerun explicitly reported agent-4f00 as its root: 2 files, 17 tests passed. Apps and embed metadata generated in agent-4f00. Direct assertion verifies sparse disabled openmates.get-docs is absent from generated Apps/settings metadata while code.get_docs remains. Structural audit rerun confirms exactly one issue: absent mindmap example. SDK/CLI parity and diff whitespace checks passed in agent-4f00.

Worktree source-path Specification check rerun passed for all three affected product-source paths.

Final scoped lint: ActiveChat contract guards, changed TypeScript lint and Svelte lint passed. Aggregate wrapper skipped tsc because it searched the tests directory for a binary, and initially failed token validation because generated tokens were missing. Generated tokens locally; reran token validator with explicit --file. Broad accidental validator invocation showed unrelated existing token errors; those are not assigned changes. No assertion of a full application typecheck or browser proof.

Explicit deletion deployment instruction

User confirms rejected-example deletion is independently deployable; replacement readiness is not a prerequisite. Missing mindmaps.mindmap coverage is an intentional temporary consequence of approved removal. Preserve the audit and record its finding; do not restore rejected content or weaken unrelated guards. Deploy cleanup and sparse OpenMates get-docs disable now, splitting get-docs if combined cleanup is gated. Registration of unreviewed replacements remains forbidden.