ai-workspace-template — Critical Review & Backlog

Open

158

Resolved

Decided

Won't fix

Strengths

Status

All 31 findings are resolved; M13 (scalar registry clobber under concurrent sessions) closed 2026-08-05 with the session-keyed registry + per-project advisory lock (TADR-0004, regression-tested). Copilot-cli measurement was live-verified 2026-08-05 (73.0k exact match), and all four design decisions settled. The initial 2026-06-24 review surfaced 12 (+ M6/M7 in follow-up); a 2026-06-25 re-review of the post-review additions — the onboard-repo workflow and the published HTML guide — added M8, L6, L7, L8, all fixed in the same pass; a further re-review of the new usage-scenarios.html added L9, also fixed. The two High items — non-functional bootstrap scripts and a non-portable Gemini hook — are fixed; the template now behaves on a clean machine as the docs describe, carries a LICENSE, ships VS Code config, checks dependencies, has an auth phase with agent runbooks, and documents a no-agent-context fallback. Each item below records what changed. New issues get appended here per the maintenance convention. A three-persona DevEx review (2026-08-11, work/devex-review/findings/) added M19–M24, L35 — seven consolidated cards covering its prioritized fix list.

Maintaining this backlog

Treat this file as a lightweight issue tracker for the template itself. Settled cards (Resolved / Decided / Won't fix) live in template-workspace-backlog-archive.html — this file holds Open cards only, so agents can read it cheaply; the scorecard counts span both files. Agents: edit both files with targeted reads (grep the ID / section), never whole-file loads. The rules:

IDs are stable and severity-scoped. High, Medium, Low, Decision + a number (e.g. M6). Never renumber or reuse an ID — once assigned it's permanent, even after the item is resolved.
Every finding carries a status badge: Open · Resolved · Decided · Won't fix.
When you fix an item: flip its badge to Resolved, add a Fixed: line describing the change (and the commit/PR if handy), and add class="resolved" to its <div class="find …">; then move the whole card to the matching severity section of template-workspace-backlog-archive.html.
When you find a new issue: append a new finding card under the right severity heading with the next free ID, status Open, and the Evidence / Impact / Fix lines.
Update the scorecard counts and the "Last updated" date in the header on every change.
Deliveries from other sessions/clones follow the same rules. A template change pushed from a downstream session must carry its backlog update in the same push; the session that pulls a change lacking one adds the missing row/card as part of incorporating it (see L14).
Commit convention: reference the ID, e.g. git commit -m "Fix M6: …", so the git log and this file stay in sync.
Re-review trigger: after any structural change to the template (new config file, new doc, new agent runtime), do a fresh evidence pass and append what it surfaces.

Tip: the matching skill for turning these into real tracked issues is to-issues (see docs/recommended-tooling.md) if you'd rather move them to GitHub Issues.

High

Medium

Low

L97 One “needs me” view across hitl nodes, BLOCKED returns and open dispatches LowOpenscripts/plan.sh, scripts/fleet.sh

Evidence — R5 (work/external-tools-review/synthesis.md, T3), accepted as later.

Impact — Low today; grows if R1 (one owner, helpers report back) brings more parallel workers.

Fix — Revisit after work/concurrent-sessions/ settles R1.

Decisions

Strengths (keep these)

✓ Secrets discipline — no tokens tracked; credentials stay in the keychain. .gitignore verified correct.
✓ Single-source agent entrypoints — one CONTEXT.md, three symlinks (mode 120000 on the remote).
✓ Layered, cross-linked docs — structure / setup / usage / tooling, one authoritative table each.
✓ All internal links resolve.
✓ Rendered onboarding — HTML guide + GitHub Pages (with .nojekyll).
✓ Accurate toolchain docs — real agent-context system + correct upstream repos.

Change log

Date	Change
2026-06-24	Initial critical review: 2 High, 5 Medium, 5 Low, 2 Decisions, 6 Strengths.
2026-06-24	Resolved all 12 findings; settled D1 (keep) and D2 (recommend `--full`). Added LICENSE, `.vscode/`, functional scripts, fallback docs, and this maintenance convention.
2026-06-24	Moved to `docs/` so it renders on GitHub Pages; referenced from `CONTEXT.md`, `README.md`, and the setup guide for agent discoverability.
2026-06-24	Added & resolved M6 (dependency-check phase) and M7 (auth phase: verify+instruct + agent runbooks under `docs/runbooks/`).
2026-06-24	Feature: added the `checkpoint` skill (`skills/checkpoint/SKILL.md` + `.claude/commands/checkpoint.md`), generalized from the Quran_Hifz_Tracker `/checkpoint` command and wired cross-agent via `CONTEXT.md`.
2026-06-25	Feature: added the `onboard-repo` workflow (`skills/onboard-repo/SKILL.md` + `/onboard-repo` + `scripts/onboard-repo.sh`) — registry identity/auth fields, graphify MCP wiring (`graphify-mcp`) with manual-scan fallback, and provenance-stamped `code-structure.md`/`design.md`/`api.md` under `docs/repo-context/<repo>/`, plus `check-repo-context.sh` freshness check.
2026-06-25	Fix: dogfooded `/onboard-repo` end-to-end against a sibling repo; `check-repo-context.sh` now resolves a repo's path from its registry `Location` (was only `repos/<name>` or root), so a repo onboarded in place at an external/local path is tracked instead of falsely reported stale.
2026-06-25	Fix: exercised the graphify live onboard path (built a real graph; `query`/`path`/`explain` verified, onboard flips to `graphify=live`). Surfaced a sibling bug — `onboard-repo.sh`'s default/`--refresh` path resolution ignored the registry `Location` (the refresh hint is pathless), so refreshing an externally-onboarded repo stamped the workspace's commit into provenance and reported `fallback` despite a live graph. Now reuses the same `Location`-first `resolve_repo_path()`.
2026-06-25	Feature: published the authoritative workspace guide (`docs/workspace-structure.md`) as a styled, self-contained `docs/workspace-structure.html` served via Pages, so humans can read it without raw Markdown. Added a reproducible generator `scripts/build-guide-html.sh` (pandoc; reuses `setup-guide.html`'s visual identity, the doc's own Contents becomes in-page nav) — generated not hand-maintained, so it can't drift from the source. Linked from the README.
2026-06-25	Re-review: fresh evidence pass over the post-review additions (the `onboard-repo` workflow + the HTML guide) per the maintenance convention. Surfaced and fixed four findings in one pass — M8 (drift guard: the guide builder embeds a source hash, `check-workspace-structure.sh` warns when the published HTML is stale), L6 (temp-file leak + non-portable `mktemp -t` in `build-guide-html.sh`), L7 (duplicated `resolve_repo_path()` extracted to `scripts/lib/repo-paths.sh`), L8 (documented the `pandoc` prerequisite). Scorecard now 0/18/2/0/6; metadata refreshed.
2026-06-25	Feature: added `docs/usage-scenarios.html` — end-to-end usage scenarios (pull template → onboard an existing repo or start a new project → build features with AI agents → checkpoint between sessions) plus a lifecycle state diagram and a feature-loop diagram. Hand-authored HTML reusing `setup-guide.html`'s visual identity with self-contained inline SVG (no pandoc/CDN dependency). Linked from the README alongside the other rendered guides.
2026-06-25	Re-review: fresh evidence pass over the new `usage-scenarios.html` per the maintenance convention — verified every command/path/skill against the repo and rendered both SVG diagrams via headless Chrome to check for collisions. Surfaced and fixed L9 (footer `../CONTEXT.md` link breaks on Pages, which serves from `docs/`; repointed to the absolute blob URL) plus a zero-impact SVG class typo. Scorecard now 0/19/2/0/6.
2026-06-28	Feature: vendored the `rlm` (Recursive Language Model) skill agent-agnostically (`skills/rlm/SKILL.md` + `scripts/rlm_repl.py`/`rlm_audit_runtime.py` + `.claude/commands/rlm.md`), sourced from brainqub3/claude_code_rlm (MIT, `LICENSE` preserved). Runtime-neutral instructions in the shared `skills/` dir; the one coupling — leaf sub-LM is a nested `claude -p` — is documented and env-overridable. Re-pointed upstream's `.claude/skills/rlm/` invocation paths to `skills/rlm/`, gitignored the REPL scratch (`.claude/rlm_state/`, `.claude/rlm_runs/`), and registered it in `CONTEXT.md` + `docs/template-usage.md`. Smoke-tested `init`/`exec` end-to-end.
2026-06-29	Feature: exported the workspace-local YouTube transcript MCP server into the template (`scripts/mcp/youtube-transcript.sh` + `scripts/mcp/youtube_transcript_mcp.py`), pre-wired it in `.mcp.json.example`, `.vscode/mcp.json.example`, and `opencode.json`, and documented the credential-free `yt-dlp`-backed setup across MCP, dependency, setup, service-access, workspace-structure, setup-guide, and usage-scenario docs.
2026-07-06	Resolved L10: folded the ask-first directive from the XDA piece "I added one line to my Claude prompts…" into principle 1 (state what you'd need to know and name silent assumptions before answering; trivial-change escape hatch), in both the `CONTEXT.md` and `docs/workspace-structure.md` copies. Wording now diverges slightly from `global.md` — sanctioned by D1. Scorecard now 0/20/2/0/6.
2026-07-14	Feature: added the `decision-log` skill (`skills/decision-log/SKILL.md` + `/decision`) — a three-tier decision-provenance scheme that captures the why code can't: Tier-1 commit trailers (always), Tier-2 ephemeral notes in `work/<proj>/decisions.md` (wide net), promoted to Tier-3 committed ADRs under a newly-shipped `docs/adr/` (`README.md` + `0000-template.md`) only for lasting-weight decisions. Promotion wired into the `checkpoint` skill (scan notes for `Promote?: yes`); ADRs + trailers registered as graphify provenance inputs so `graphify query "why…"` can walk `code → commit → ADR → alternatives-rejected`. Registered in `CONTEXT.md` (new Decision Records section + skills entry); narrowed D2 (template now ships `docs/adr/` itself, not just via `init --full`).
2026-07-22	Fix M9: first live use of the D4 context-budget system in a fresh session caught `register` trusting the stale session registry instead of re-discovering — bound the previous session's transcript and reported a false WARN. `resolve_session()` now forces fresh discovery for `register` only. Scorecard now 0/21/4/0/6.
2026-07-22	Feature (D4 follow-up): enabled Gemini CLI local telemetry for exact context counts — tracked `.gemini/settings.json` gains a `telemetry` block (`target: local`, `logPrompts: false`, outfile `.gemini/telemetry.log`, gitignored); the gemini adapter in `context-budget.sh` prefers the workspace telemetry log and parses the last response's input tokens (both `input_token_count` and OTel `gen_ai.usage.input_tokens` spellings), estimating from the chat log — never the telemetry file — when no response exists yet. Wiring proven live (a real run wrote the log); parser fixture-verified for both spellings; a successful-response verification is blocked on Gemini auth. Copilot CLI follow-up recorded as blocked: the CLI isn't installed on the origin machine.
2026-07-23	Fix M10: a live copilot-vscode session (downstream workspace) caught `copilot_vscode_discover()` pinning a stale sibling session's log by mtime and reporting a false STOP — the discover now prefers `$VSCODE_TARGET_SESSION_LOG` (the live log's path, exported by Copilot terminal sessions), falling back to the newest-mtime scan.
2026-07-23	Session-pinning review across all adapters (same bug class as M10: newest-mtime as a proxy for "current session"). Fixed M11 — `claude_discover()` now pins `$CLAUDE_CODE_SESSION_ID.jsonl` (differential-verified: with a stale transcript touched newest, unpinned `register` binds it, pinned binds the live session). Opened M12 (gemini shared telemetry log serves the previous session's counts to a new session's `register`), L11 (codex/copilot-cli mtime races; codex any-project fallback), L12 (registry keyed per-runtime clobbers under concurrent same-runtime sessions). Scorecard now 3/23/4/0/6.
2026-07-23	Landed the three non-gated session-pinning follow-ups. M12 resolved — `register --runtime gemini` truncates the shared workspace telemetry log after binding it, and `gemini_measure` reports `0 estimate` for a fresh session with no logs at all; lifecycle fixture-verified (stale 140K WARN → 0 after register → appended response reads exact). L11 partially landed — codex any-project fallback removed (fails over binding another project's rollout; verified against 4 non-matching rollouts); the env-var pin checks stay open, machine-gated. L12 resolved — single-session-per-runtime-per-workspace documented as an explicit limitation in `docs/context-budget.md`. Scorecard now 1/25/4/0/6.
2026-07-23	Opened L13 (surfaced while walking through the register lifecycle): session registration is documented for agents only — the developer quickstart never mentions it, and no `SessionStart` hook makes it mechanical. Fix queued for the next session. Scorecard now 2/25/4/0/6.
2026-07-23	M10 follow-up (`7c29476`, delivered from the origin copilot workspace and pulled into the template): `$VSCODE_TARGET_SESSION_LOG` on current Copilot builds points at a `debug-logs/<id>` directory rather than the token file — `copilot_vscode_discover()` now derives the session id from the var's basename and pins `chatSessions/<id>.jsonl`, keeping discovery deterministic; newest-mtime survives only as the var-unset fallback. Adapter table updated. No scorecard change (M10 already resolved).
2026-07-23	Resolved L13: session-lifecycle paragraph added to the developer quickstart in `docs/context-budget.md` (registration is the agent's job — automatic-by-instruction via `CONTEXT.md`, so only for sessions started inside the workspace tree by an instruction-following agent; unregistered sessions fall back to newest-mtime discovery), plus a `SessionStart` hook in `.claude/settings.json.example` that runs `register --runtime claude --transcript <path from the hook payload>`, making registration mechanical for Claude Code. Hook event live-verified in the resolving session; command verified against a simulated payload (pins the exact transcript) and fails open on an empty one. Scorecard now 1/26/4/0/6.
2026-07-23	Resolved the L11 remainder — the machine-gated session-pin env-var checks. Codex now pins `rollout-*-$CODEX_THREAD_ID.jsonl`, live-verified on the origin machine (env var probed inside a real `codex exec` session equals the rollout UUID and `payload.id`, injection confirmed in `openai/codex` source; differential-verified against a stale rollout touched newest). Copilot CLI now pins `session-state/$COPILOT_AGENT_SESSION_ID/events.jsonl` under `${COPILOT_HOME:-~/.copilot}` and scans the correct `session-state/` path (dropped the unattested `sessions/` path, kept legacy `history-session-state/`) — research-sourced from the CLI changelog and fixture-verified, but live verification still pends an install so the adapter stays unverified. Scorecard now 0/27/4/0/6.
2026-07-23	Feature (`bd40023`, delivered from a downstream session and pulled into the template): formalized the work-directory launcher/ledger convention — forward `next-session.md` (REPLACED each rollover) + append-only `handoff.md` (newest block on top, archived to `handoff-archive.md`) + required `README.md` — in new `docs/work-directory-conventions.md`, with a `create-work-item` skill (`/create-work-item`) and wiring in `CONTEXT.md`/`workspace-structure`. On pull, `work/context-decay/` was migrated to the convention: README added, purpose headers stamped on both backbone files, and the ledger rebuilt with the two most recent session blocks (session 3's recovered from git history). No scorecard change.
2026-07-23	Feature (`f3928e6`, delivered from a downstream session and pulled into the template): `scripts/diff-review.sh` — symlink-safe commit/range directory-diff review. Always passes `--no-symlinks --no-prompt` (git's dir-diff temp trees use symlinks for the working-tree side, which Beyond Compare renders as missing files) and prefers the blocking `bcomp` launcher on macOS (the non-blocking one races git's temp-dir cleanup → empty panes); VS Code Compare Folders / per-file fallbacks via `-t`. Documented in `docs/operational-knowledge.md` ("Diff Review Workflow") and `workspace-structure`. On pull: syntax + `--help` verified, doc references resolve. No scorecard change.
2026-07-23	Opened + resolved L14: three downstream-delivered template changes in one day (`7c29476`, `bd40023`, `f3928e6`) all skipped the backlog convention and needed after-the-fact reconciliation by the pulling session. The convention now covers both ends explicitly (new "Maintaining" rule + `CONTEXT.md` Template Backlog sentence). Also slimmed the user's global agent context: the manual Beyond Compare persistent-dirs recipe replaced with a pointer to `scripts/diff-review.sh`. Scorecard now 0/28/4/0/6.
2026-07-24	Opened + resolved L15: audited the installed skill set against `recommended-tooling.md` after reviewing upstream `mattpocock/skills` (local clone at `ed37663` = upstream HEAD; no new released skills since the Jul 22 sync). Added the two undocumented `misc`-bucket skills (`git-guardrails-claude-code`, `setup-pre-commit`), a "skill not found on another machine" staleness/recovery note (prompted by a real `teach` miss on a second machine whose clone predated the skill), and a caveat on upstream's new native Claude Code plugin distribution — Claude-Code-only, incompatible with the agent-agnostic symlink install. Scorecard now 0/29/4/0/6.
2026-07-30	Feature: wayfinder integration (hybrid, per `docs/superpowers/specs/2026-07-30-wayfinder-integration-design.md`). Vendored upstream `wayfinder` into `skills/wayfinder/` (from `mattpocock/skills` @ `2ab9580`, provenance comment with pinned commit + refresh procedure) so Copilot users and template downloaders get it agent-agnostically; new `docs/agents/issue-tracker.md` — local-markdown tracker adapted from upstream's `.scratch/` to `work/<effort>/` (map at `map.md`, tickets under `issues/`), with a decision-log tie-in (resolved ticket = Tier-2 note). Wired: `CONTEXT.md` Workspace Skills bullet, `/wayfinder` shortcut, `workspace-structure` + `work-directory-conventions` registration (also fixed `create-work-item` missing from the structure doc's skills listing). `recommended-tooling.md` refreshed after upstream sync (clone at `2ab9580`): bucket list gains `in-progress/`, vendored-copy/duplicate caveat, and a "newer upstream skills worth watching" note (`to-questionnaire`, `wizard` recommended; `batch-grill-me` nice-to-have; `claude-handoff` anti-recommended vs `session-rollover`). No scorecard change.
2026-08-05	Upstream skill sync (`mattpocock/skills` clone `2ab9580` → `8b36d4f`). Vendored `skills/wayfinder/` refreshed to `8b36d4f` (one substantive upstream line — the Grilling ticket type now always invokes `/grilling` + `/domain-modeling`; `agents/openai.yaml` unchanged; diff vs upstream re-verified comment-only). Global symlinks: removed broken `writing-great-skills` (upstream renamed/restructured it to `writing-for-agents`), linked the four newly released skills — `writing-for-agents`, `wizard`, `to-questionnaire`, `wait-what`; `global.md` setup/util list updated to the new name. `recommended-tooling.md` §3: table gains `wizard`/`to-questionnaire`/`wait-what` rows, `writing-great-skills` renamed; "worth watching" note rewritten — graduated skills dropped, `batch-grill-me` noted as folded into the `grilling` engine, still-in-progress list now `claude-handoff` (anti-recommended), `loop-me`, `setup-ts-deep-modules`, `writing-*`. No scorecard change.
2026-08-05	Feature: skill-hardening batch — 8 improvements to the workspace skills from a full comparison against `mattpocock/skills` @ `8b36d4f` (plan: `work/template-maintenance/skill-hardening-plan.md`). (1) `checkpoint` promotion scan extended to `work/*/map.md` Decisions-so-far (wayfinder decisions no longer escape promotion; cross-ref in `decision-log`). (2) Placement/consistency: `onboard-repo`'s mandatory budget cadence moved above its steps; convention pointers unified on `CONTEXT.md` (`rlm`, `issue-tracker.md`); `create-work-item` optional files now list `spec.md`/`map.md`/`issues/`. (3) Runnable Verification sections added to `checkpoint`, `create-work-item`, `session-rollover` (newest-on-top + archive checks), `decision-log`. (4) Identical first-yes-wins "Which boundary skill?" block in `checkpoint` + `session-rollover` (WARN/STOP → rollover; measurement wins). (5) Dangling global-`handoff` prerequisite resolved by inlining its three load-bearing rules into both boundary skills (Tier-2 note). (6) `disable-model-invocation: true` on the user-driven `create-work-item`/`onboard-repo`/`rlm`. (7) ADR bar restated as the three-way AND test (hard-to-reverse ∧ surprising ∧ real trade-off, from upstream `domain-modeling`) in `decision-log` + `docs/adr/README.md`, with a worked "don't log this" Tier-2 counter-example. (8) Vendored `skills/writing-for-agents/` (SKILL.md + SKILL-MECHANICS.md + `agents/openai.yaml`, pinned `8b36d4f`, diff-vs-upstream comment-only) and wired it into `CONTEXT.md` + `recommended-tooling.md` §3 (vendored-skills blockquote now covers both copies). No scorecard change.
2026-08-05	Feature: rollover relaunch script — `scripts/launch-next-session.sh` (TADR-0003/0004 item #2): launches a successor session seeded with the canonical bootstrap prompt (verbatim, single source of truth in the script); runtime from the dying session's own registry record (D6), `ROLLOVER_RUNTIME` fallback-only; 5 runtimes seeded-interactive (`claude`, `codex`, `gemini -i`, `opencode --prompt`, `copilot -i`; flags re-verified against live `--help`), `--bg` claude-only with D8 successor confirmation; `ROLLOVER_RELAUNCH=off\|manual\|auto` honored; non-tty manual prints the ready-to-run command. Tests: `scripts/tests/test-launch-next-session.sh` (13 cases, dry-run + stub-binary launch paths). Status note flipped in `docs/context-budget.md`. No scorecard change.
2026-08-06	Feature: vendor hook deployments (TADR-0003/0004 item #3) — context-budget WARN/STOP now pushes in-band to all five runtimes, not just Claude Code. Extracted the shared throttle/escalation/fail-open core into `scripts/hooks/context-budget-hook-lib.sh` (`4a39bf8`), then thin per-runtime wrappers: opencode sqlite measurement adapter (`6a78138`), codex `UserPromptSubmit` hook via hash-trusted `.codex/config.toml` (`4543645`), gemini `BeforeAgent` hook reading the workspace telemetry log (`93e9d45`), opencode `chat.message` plugin registered in `opencode.json` (`f9579fe`), and copilot CLI `sessionStart`/`agentStop` hooks gated by `~/.copilot/config.json` `trustedFolders`, blocking only at STOP (`9fa16b7`). Full per-runtime wiring/channel/friction table in `docs/context-budget.md` → "Vendor hook deployments". No scorecard change.
2026-08-06	Feature: successor option inheritance (TADR-0003/0004) — `e0cb70d`: `scripts/launch-next-session.sh` now replays the dying session's launch options (normalized approval level, model, raw-extra escape hatch) onto the successor, persisted via `work/<project>/.rollover-options` (written by `session-rollover` step 6) and mapped to each runtime's own flags. Flag-verification correction: the plan's table guessed codex `--full-auto`, which does not exist in codex-cli 0.142.4 — shipped `--ask-for-approval never` instead (live `--help`-verified); opencode's `--auto` flag does exist, so opencode stayed in the approval mapping (the plan's drop-contingency wasn't needed). Documented in `docs/context-budget.md` → "Relaunch knobs". No scorecard change.
2026-08-06	Feature: attach-session.sh helper (Task 9, user-added to the vendor-hook-deployments plan — overrides the earlier YAGNI call to leave re-attach as documented-only) — `3fafd13`: `scripts/attach-session.sh <project>` resolves the latest session for a work item (`.active-session` lock, falling back to the newest `.context-budget/sessions/` record for the project), prints a one-line status (`project=… runtime=… session=… age=…s live=… locked=…`), and when it is alive and locked, connects this terminal to it. Fixed a latent doc bug along the way: `docs/context-budget.md`'s "Chained rollovers & re-attach" passage named a `claude attach` subcommand that does not exist — re-verified live against `claude --help`/`claude agents --json` 2026-08-06, no `attach` subcommand, `-r/--resume <session_id>` is the closest supported form and is what the script wires (`run: <cmd>` in `--dry-run`/non-TTY, same idiom as `launch-next-session.sh`). Non-claude runtimes get a "cannot attach" message (no background sessions to attach to) instead of guessing a flag. Tests: `scripts/tests/test-attach-session.sh` (6 cases, liveness driven by artifact mtime vs `CONTEXT_LOCK_STALE_SECS`). Decision + rejected alternatives in `work/automatic-session-rollover/decisions.md`. No scorecard change.
2026-08-06	Final whole-branch review of vendor-hook-deployments. Fixed the last stale `claude attach` reference: `skills/session-rollover/SKILL.md`'s closing paragraph pointed at the nonexistent `claude attach` subcommand (same bug already fixed in `docs/context-budget.md` by `3fafd13`, missed here) — repointed to `scripts/attach-session.sh <project>`. Opened L17: bundled four minor issues the review triaged as non-blocking follow-up (attach-session.sh's contradictory "live but unlocked" wording, a space-unsafe `ls -t` glob loop shared by `attach-session.sh` and `launch-next-session.sh`, an unquoted SQL interpolation in `context-budget.sh`'s `opencode_measure`, and a stale test-suite filename in `context-budget.md`) so they aren't lost. Scorecard now 2/29/4/0/6.
2026-08-06	Feature: per-work-item relaunch override — `8817d1d`: an optional, committed `work/<project>/context-budget.env` now overrides the workspace-global `ROLLOVER_RELAUNCH`/`ROLLOVER_RUNTIME` for that work item; `scripts/launch-next-session.sh` sources it after the global env (sourcing moved below arg parsing, since the per-item path needs `$PROJECT`), with precedence explicit env var > per-item file > global `context-budget.env` > built-in default (`off`). Only the relaunch knobs are per-item — WARN/STOP thresholds stay workspace-global — but the file name/shape matches the global env so thresholds could ride in later without a rename. Rejected alternative (knob in `.rollover-options` — per-launch provenance, would be clobbered at each rollover) recorded in `work/per-item-relaunch-override/decisions.md`. Tests T15a–d added to `test-launch-next-session.sh` (42 asserts green); docs updated (`context-budget.md` "Relaunch knobs" — the old "no per-project override" claim replaced — plus `CONTEXT.md` and the `session-rollover` skill). Motivating case applied: `work/automatic-session-rollover/` now carries `ROLLOVER_RELAUNCH=auto`. No scorecard change.
2026-08-06	Improvement: learnings capture across rollovers — three-rule scheme codified after a user retrospective discussion: (1) route learnings at incident time to their existing durable homes (`operational-knowledge.md`, `decisions.md`, docs), making the rollover Reflect step a sweep for what slipped; (2) observations not obviously durable get parked as one-line `Learnings:` entries in the handoff ledger block; (3) second-strike promotion — a parked learning graduates to a durable home only when it bites a second time (grep `handoff*.md` for the prior strike), so single events die in the archive. Edits: `skills/session-rollover/SKILL.md` (Reflect step + handoff contract) and `skills/checkpoint/SKILL.md` (human-driven retro belongs at checkpoint or successor-session start, never in a token-starved rollover). Deliberately no new `learnings.md` file and no auto-retro machinery — the bottleneck is filtering, not capture. No scorecard change.
2026-08-06	Opened + resolved M14 (rollover launched the successor while the predecessor still held the work-item lock — fired live on this work item's first `ROLLOVER_RELAUNCH=auto` relaunch; the successor's `register` was refused and `LOCK_STALE=3h` gave no reclaim path). Fix: `launch-next-session.sh` now releases the dying session's own lock before any launch path (identity-matched against its registry record; foreign locks never removed; `--dry-run` exempt; `mode=off` included), replacing the skill's launch-then-release ordering, which could never work in auto mode (register races the lock) or attached-manual mode (`exec` forecloses the release step). Skill + `context-budget.md` reworded to release-before-launch with the manual fallback retained; tests T16–T19 added to `test-launch-next-session.sh` (49 asserts green). Opened L18: `context-budget.sh`'s env-file sourcing guard keys on `CONTEXT_DUMB_ZONE_TOKENS` only, so explicit overrides of other knobs (hit: `CONTEXT_LOCK_STALE_SECS`) are clobbered by the sourced file. Scorecard now 3/30/4/0/6.
2026-08-06	Improvement: runtime state untracked — architecture review (one dev, one machine: commit what a future session must read; ignore live-session state whose validity is artifact mtime, not content). Untracked `work//.active-session` (a committed lock is a stale claim resurrected on checkout; post-M14 the launcher deletes/recreates it every rollover) and `work/context-decay/context-ledger.jsonl` (hook-appended telemetry, same class as the already-ignored `.gemini/telemetry.log`; not regenerable, so analysis snapshots get committed deliberately — as `subagent-rollover-stats.md` did — never the live file); pre-emptively ignored `work//.rollover-options` (per-launch machine-local flags, per the per-item-relaunch-override decision). Rationale codified in `.gitignore` comments, `docs/context-budget.md` (Ledger + work-item-ownership passages), and `docs/work-directory-conventions.md` (new "Tracked vs. untracked" rule). Side effect: the perpetually-dirty tree is gone, so clean-branch checks at checkpoint/rollover are meaningful again. No scorecard change.
2026-08-06	Opened + resolved L19 (no normalized approval level mapped to claude's classifier-vetted auto mode; the script's `auto` collided with the UI's "auto mode" while actually meaning acceptEdits — raised as `work/automatic-session-rollover/issues/02` after the first auto-relaunched successor came up in manual mode). Fix: `ROLLOVER_OPT_APPROVAL` ladder is now `default < edits < auto < full` — `auto` re-pointed to the classifier tier (claude `--permission-mode auto`, verified against live `--help` 2.1.223 per TADR-0003; non-classifier runtimes fall back to their `edits` mapping with a stderr note), `edits` carries the old acceptEdits-tier mappings. Tests T14a–T14p reworked/added in `test-launch-next-session.sh` (55 asserts green); `context-budget.md` "Option inheritance" + `session-rollover` skill step 6 updated; alternatives (a claude-centric `classifier` level name; retiring `auto` loudly, which degrades to manual mid-chain) recorded in the work item's `decisions.md`. Scorecard now 3/31/4/0/6.
2026-08-06	Opened + resolved L20 (plainly-exited sessions squatted their work-item lock for the whole 3h stale window — the M14 failure mode, mitigated only for the rollover path — and nothing distinguished a live helper session from a dead predecessor). Fix: session-role model `primary / auxiliary / superseded` with the lock as the authoritative primary marker; `register` records + prints `role=`, the launcher stamps `superseded` at pre-launch release, `attach-session.sh` shows `role=`, a `SessionEnd` hook releases the lock on plain exit, and a new Claude Code statusline script (`scripts/statusline-context-budget.sh`) shows `PRIMARY · <project> · <pct>%`. One-primary contract documented in `context-budget.md` "Session roles" + `work-directory-conventions.md`; deferred refinements (superseded_by back-link, --takeover, sessions listing, tab-title display) in `work/automatic-session-rollover/issues/03-session-roles.md`. Tests: registry T6, launcher T20, attach T7, new statusline suite — 107 asserts green across the four touched suites. Scorecard now 3/32/4/0/6.
2026-08-06	Improvement: meaningful session titles for rollover successors (user finding: a successor's auto-generated title read "session lock verification" — the title generator latched onto the launcher's first mechanical step). `launch-next-session.sh` now leads the bootstrap prompt with `Work item <project> - rollover session #N` (helps every runtime's title generation) and passes claude `--name "<project> #N"` (`-n/--name` verified against live `--help` 2.1.223: session display name for picker + terminal title; no equivalent verified for other runtimes). N tracked in machine-local `work/<proj>/.session-seq` (gitignored; incremented per real launch, never by `--dry-run`). Gotcha fixed en route: the prompt must stay ASCII — an em-dash in it broke the `%q`-escaped `cmd:` echo through BSD sed ("illegal byte sequence"). Tests T21a–g (launcher suite 65 green); `session-rollover` skill step 7 + `context-budget.md` "Relaunch knobs" updated. No scorecard change.
2026-08-06	Opened + resolved L21 (the L20 project-level `statusLine` replaced the user's global status bar — `ccstatusline`'s model/context/cost/git widgets vanished — and unregistered sessions saw a bare `no work item`; user report). Fix: `statusline-context-budget.sh` chains the global `statusLine.command` from `~/.claude/settings.json` (same stdin, output first, budget segment appended as an extra line only when the session is on a work item; recursion-guarded). Tests T7a–f, statusline suite 14 asserts green. Scorecard now 3/33/4/0/6.
2026-08-06	Improvement: rollover implementation slice 1 — parent/child session registry + lock lineage (sessions 16–17; closes the deferred items in `work/automatic-session-rollover/issues/03`; no new finding card — feature work tracked in the work item, recorded here per the delivering-commit rule). Role set extends to `child`: `register --parent-session <sid> [--agent-id]` registers a subagent parent-side with artifact-derived identity (`parent_session_id`/`depth`/`agent_id`); per-child locks in `work/<p>/.agent-locks/` granted only when the parent chain terminates at the project-lock holder; `release` sweeps stale child locks then refuses bottom-up violations (I4). Rollover lineage now closes both ways: successor back-stamps `superseded_by` at primary acquisition (T13), and `register --takeover` is an explicit recorded steal — wins against a live holder, stamps the loser's record (T14). Registry suite T7–T14 (54 asserts), all five suites green (192 total). Docs: `context-budget.md` multi-session model reworked to four roles; decisions promoted to TADR-0005. No scorecard change.
2026-08-06	Improvement: rollover implementation slice 2 — per-child transcript sweep (R1) (session 18; feature work tracked in the work item, recorded here per the delivering-commit rule). New `context-budget.sh children [--parent-session <sid>] [--all]`: enumerates a Claude parent's `subagents/agent-.jsonl` + `.meta.json`, measures each with a sidechain-inclusive* Claude-measure variant (a subagent's rows are all `isSidechain:true` — the self-measure filter would silently degrade every child to a size estimate), prints escalation-only WARN/STOP lines with `age=`/`type=`, exits with the worst child status; non-claude runtimes die loudly. New suite `test-children-sweep.sh` C1–C9 (27 asserts); all six suites green (219 total). Live smoke test surfaced the research's motivating 141.8K subagent as WARN. Docs: `context-budget.md` "Per-child sweep" section + quickstart line. No scorecard change.
2026-08-06	Improvement: workspace-root anchoring — coordination state keyed to repository, not checkout (session 19; resolves `work/automatic-session-rollover/issues/05`; feature work tracked in the work item, recorded here per the delivering-commit rule). All five coordination scripts (`context-budget.sh`, `launch-next-session.sh`, `attach-session.sh`, `statusline-context-budget.sh`, `context-budget-hook-lib.sh`) resolve `WORKSPACE_ROOT` via `git rev-parse --git-common-dir`, so every git worktree converges on the main checkout's `.context-budget/`, locks, ledger, and `.session-seq`; script-relative fallback outside git. `launch-next-session.sh` invoked from a worktree now syncs the main checkout (pushed-check, ff-only pull, loud dirty/diverged refusals) and launches from the main root — retiring the "no auto-relaunch from worktrees" ban and the register-before-isolate discipline. New tests G1/G2 (registry), W1–W5 (launcher), T8 (statusline); all six suites green (243 total). Docs: `context-budget.md` "Worktrees" section; operational-knowledge divergence entry marked superseded-by-fix. No scorecard change.
2026-08-06	Improvement: session-19 decision note promoted to TADR-0006 (session 21) — Key coordination state to repository identity, never to a checkout; amends TADR-0004/0005's implicit one-checkout assumption. No scorecard change.
2026-08-06	Improvement: rollover implementation slice 3 — dispatch-contract hardening (R2/R3) (session 21; feature work tracked in the work item, recorded here per the delivering-commit rule). New stateless, runtime-agnostic `context-budget.sh dispatch-contract --report <path> [--brief <path>] [--gen <n>]`: emits the R2 rollover contract for a long-running child's dispatch prompt (checkpoint-at-boundaries to the report file, 15-line return cap, `DONE \| DONE_WITH_CONCERNS \| BLOCKED \| NEEDS_CONTEXT \| ROLLOVER_NEEDED` vocabulary, ROLLOVER_NEEDED only-on-request, gen≥2 read-report-first clause; ASCII-only for the `%q`/BSD-sed launch paths). R3 documented as parent-side policy: successor dispatch is the only rollover verb, resumes stack history — sweep (`children`) before resuming, roll at WARN/STOP. New suite `test-dispatch-contract.sh` K1–K7 (24 asserts); all seven suites green (267 total). Docs: `context-budget.md` "Dispatching long-running children" section + quickstart line; CONTEXT.md dispatch-time pointer. No scorecard change.
2026-08-06	Improvement: rollover implementation — parent-persisted dispatch records + generation fencing (R4) (session 22; feature work tracked in the work item, recorded here per the delivering-commit rule). Three new `context-budget.sh` subcommands: `dispatch-open --project <p> --task <slug> --report <path>` appends generation N+1 to the gitignored record `work/<p>/.agent-dispatch/<slug>.json` (agent-type/model/effort/brief/agent-id persisted for faithful re-dispatch) and emits the R2 contract for that generation in one step, refusing while the previous generation is still open (fencing: at most one live writer per report file); `dispatch-close --status <five yield statuses\|KILLED>` closes it; `dispatch-list` exits 1 while any generation is open (the rolling parent's drain check). Contract emitter now labels progress blocks `[gen N]`. This is what makes parent rollover fleet-safe: a successor parent reconstructs the orchestration from records + reports and re-dispatches unfinished subtrees fresh. New suite `test-dispatch-records.sh` L1–L9 (51 asserts); all eight suites green (318 total). Docs: `context-budget.md` section now R2–R4 + quickstart bullet; CONTEXT.md dispatch bullet; `.gitignore` runtime-state entry. No scorecard change.
2026-08-06	Improvement: rollover implementation — register-time sweep of stale primary records (registry hygiene) (session 23; feature work tracked in the work item, recorded here per the delivering-commit rule). Primary acquisition in `context-budget.sh register` now stamps any other same-project `role=primary` record whose artifact liveness is stale (same `LOCK_STALE` rule as the lock) with the takeover triple `role=superseded` + `superseded_at` + `superseded_by`; live records, other projects' records, and non-primary roles are untouched, and auxiliary/child registrations never sweep. Retires the parked sessions-19/21 learning (records from sessions that died without release lingered until the 7-day GC). Registry suite T15 (8 asserts, 68 total); all eight suites green (326 total). Docs: one paragraph in `context-budget.md` roles section. No scorecard change.
2026-08-06	Finding: in-band WARN push is swallowed when a busy child wins the throttle race (claude runtime) (session 24, ticket 06 empirical run). PostToolUse hooks fire on a subagent's tool calls keyed to the PARENT `session_id` (shared throttle stamp + shared escalation status), but the harness drops hook exit-2 stderr for those fires: when the 120K WARN crossed mid-run, the one-shot escalation was consumed on a child fire and the message reached no transcript (parent, child, or grandchild) — escalation-only then never re-emits, so a session running a high-cadence child effectively loses its WARN push. Mitigation candidates: skip state-write/emit when hook stdin identifies a child fire; or treat `record`-driven measurement as the WARN source of truth while `dispatch-list` shows open generations. Evidence: `work/automatic-session-rollover/research/06-midflight-hook-injection.md`. Status: Open.
2026-08-06	Finding: EnterWorktree mid-session relocates the claude transcript; the registry artifact path goes stale (session 24). Entering a worktree moves the session JSONL to the worktree's project slug (`…--claude-worktrees-<name>/<session>.jsonl`); the `.context-budget/sessions/` record registered pre-worktree keeps the old path, so registry-artifact-based liveness/token checks (lock staleness, sweep, `check` without `--transcript`) misread the session as dead/stale. Hook-side checks are immune (stdin `transcript_path` is current). Fix directions: re-register after EnterWorktree, or resolve the artifact by session-id glob at check time. Second strike 2026-08-06 (session 25): bit a live lock decision, not just a check — session 25 registered as primary at 20:27, entered its worktree at ~20:28 (transcript relocated, registry path dead), and at 20:30 a newly-registering session could not confirm its liveness (`lock_holder_age` rc 1 = treated-as-stale), swept it, and took primary while session 25 was mid-work. Two live-by-the-rules takeovers of a live session in two sessions — promote the fix (re-register after EnterWorktree is the cheap half; session-id glob resolution the robust half). Status: Open.
2026-08-06	Design: copilot measured-tier adapter — designed, build unscheduled (session 25, wayfinder ticket 09; decision recorded in the work item, this row per the delivering-commit rule). The adapter is an in-place extension of `context-budget.sh`, no new script: composite child identity `<parentSessionId>+<toolCallId>` carried by the existing `--agent-id` flag (copilot `task` children have no session of their own); `children` grows a copilot-cli branch (discovery via `subagent.started` events cross-checked against `session-store.db::assistant_usage_events initiator='sub-agent'`); measurement keeps claude semantics — last usage row's `input+cache_read+cache_write` per agent_id (`method=exact`, WAL-sidecar snapshot before reading), `subagent.completed.totalTokens` as estimate-only fallback (lifetime sum, not context size); per-child locks and R4 dispatch records reused verbatim; escalation is post-hoc/external-observer only, since a sync `task` blocks the parent and ticket 06 closed the push-injection tier. Build slices (incl. slice-0 probe of `mode: background`/`write_agent` stream shape, and a sqlite-first upgrade of `copilot_cli_measure`, whose current grep silently size-estimates) await user scheduling. Full design + rejected alternatives: `work/automatic-session-rollover/issues/09-copilot-adapter-design.md`. No scorecard change. Session-26 verification: child events in the durable evidence files carry only `outputTokens`/`totalTokens` (every `inputTokens` sits on the single `session.shutdown` rollup) — sqlite-first is confirmed necessary, and the `copilot_cli_measure` silent-degrade concern is confirmed real (nothing greppable exists mid-session).
2026-08-06	Finding: launcher (`next-session.md`) staleness race — successor session trusted a stale launcher because the local checkout lagged origin/main (session 26). Session 25 resolved wayfinder ticket 09, REPLACED the launcher, and pushed to origin/main from its worktree; the local main checkout was never ff-pulled, so session 26 launched against session 24's launcher, whose freshness first-action ("confirm the checkout carries the session-24 commit, pull if not") passed vacuously — it keys on the previous session's commit marker and cannot detect a newer sibling rollover that exists only on the remote. Session 26 re-resolved ticket 09 in parallel before the push rejection exposed the race; reconciled with origin canonical (duplicate kept on branch `session-26-pre-reconcile`). Mitigation shipped in the launcher: First-actions now require `git fetch origin` + empty `git log HEAD..origin/main` BEFORE trusting the launcher's content. Broader fix direction (unbuilt): `launch-next-session.sh` / the work-directory convention should guarantee the ff-pull mechanically before a successor reads `next-session.md`, instead of relying on a prose first-action. Status: Open.
2026-08-06	Fix M10 second follow-up (session 27): `copilot_vscode_discover()` failed under the sandboxed VS Code terminal — the sandbox blocks `readdir` on the `workspaceStorage/` parent, so the glob-and-grep workspace scan silently expanded to nothing even though the token file itself stays readable via a direct path. Discovery now derives the workspace-storage hash from `$VSCODE_TARGET_SESSION_LOG` by parameter expansion (the debug-logs path embeds it) and probes `chatSessions/<sid>.jsonl` directly before falling back to the glob scan (older builds / var unset unchanged). Verified with a fake-`HOME` harness including a `chmod 311` readdir-blocked simulation (7/7), all eight `scripts/tests/` suites green (326 asserts), and a live in-copilot run of the spec's Verify section (sandboxed terminal, user-relayed): correct artifact pinned with no directory listing; `method=estimate` mid-first-turn is the designed pre-usage-flush degrade — the same file measured `38152 exact` once the turn flushed. Spec (user-approved, session 26): `work/context-decay/copilot-vscode-sandbox-discovery-fix.md`. No scorecard change (M10 already resolved).
2026-08-06	Ship Copilot VS Code agent-mode context-budget wiring (issue 01, sessions 28–29): new `scripts/hooks/context-budget-copilot-vscode-hook.sh` + committed `.github/hooks/context-budget-vscode.json` (PascalCase events, repo-relative command — hook-process cwd verified as the workspace root, VS Code 1.132.0). SessionStart pushes WARN/STOP as `hookSpecificOutput.additionalContext`; Stop blocks ONLY at STOP via stderr + exit 2 (JSON `decision:block` verified IGNORED by VS Code), guarded by `stop_hook_active`. Runtime discriminator: VS Code payloads are snake_case (`session_id`) vs Copilot CLI camelCase (`sessionId`) — both `.github/hooks/*.json` files coexist safely. Token artifact derived hook-side from `transcript_path` → `chatSessions/<sid>.jsonl` (promptTokens re-confirmed live). `launch-next-session.sh` gains a real copilot-vscode seeded launch: `code chat -r -m agent "<prompt>"` (verified: opens a new agent session in the last-active window, exits before it responds → always routed through the BG confirm loop). Tests: vendor T11 block (12 asserts, fake workspaceStorage tree) + launcher T22 (argv + implied bg). Docs: `context-budget.md` (runtime table hook route, vendor deployments row, relaunch knobs), `relaunch-analysis.md`, CONTEXT.md five→six runtimes.
2026-08-06	Session-number single source of truth (wayfinder issue 10, session 30; TADR-0007): the three number carriers (machine-local `.session-seq`, `handoff.md` ledger titles, worktree names) had drifted — a session launched as prompt-#28 was ledger-#29, from a hand-pasted (launcher-bypassing) launch that skipped the seq increment. Rule shipped: `.session-seq` + the bootstrap prompt are canonical; sessions use their prompt number verbatim in ledger titles and worktree names, ledger repaired on disagreement; drift self-heals via a new `session-rollover` step-7 sync (the dying session writes its own number to the counter before launching). Rejected: ledger-prose parsing as source or as a `register`-time checker (fragile / YAGNI), committing the counter (multi-machine conflicts). Docs: `work-directory-conventions.md` (ledger section), `context-budget.md`, `skills/session-rollover/SKILL.md`. No launcher code change; no scorecard change.
2026-08-07	Resolved L17 + L18, the last two Open cards. L17 (four deferred rollover-script issues): attach-session live-but-unlocked message now matches its own flags; both `ls -t` glob loops space-safe via `while IFS= read -r`; `opencode_measure` SQL-escapes `$PWD`/`$sid`; the stale registry-suite filename survived only in the test file's own `# File:` header (the `docs/context-budget.md` reference was already correct), now fixed. L18: per-variable precedence around the `context-budget.env` source in `context-budget.sh` (capture-before/restore-after, the `launch-next-session.sh` pattern) — default-only env-file assignments rejected (would invert the per-item override chain). New regression test T16 (red on pre-fix script, green after); T4d updated to the new attach wording. All eight suites green (343 asserts). Scorecard now 0/46/4/0/6.
2026-08-07	Improvement: scripts/context-inspect.sh — context composition analyzer for the context-decay work item (user request). Exact per-turn totals from the usage envelope; harness-attachment breakdown split turn-1 vs later (itemizes the L29 ~10K first-message materialization: skill_listing ~4K, deferred_tools ~2.4K, hook context ~2K, mcp_instructions ~0.9K, agent_listing ~0.8K); disk-side CLAUDE.md/memory stack (harness-injected, never transcribed); harness-fixed remainder. Verified on two live transcripts. Doc pointer in `context-budget.md` → Quickstart — agent. Claude-runtime transcripts only. No scorecard change.
2026-08-07	Resolved L31 (session #4, settling the attribution question parked on the first L19 card): residual analysis over two Warp-plugin-heavy transcripts proved `hook_success` stdout never enters model context — only the attachment's `content` field does. `context-inspect.sh` now attributes hook_success at content length (was whole-JSON, over-attributing ~20–29K tok on Warp-heavy sessions). Same session: turn-1 trim candidates analyzed and declined by user decision (no repo-side item is both unused and consequence-free; skill_listing is only ~307 tok workspace-controlled) — see `work/context-decay/trim-estimates.md` + decision note. Scorecard now 0/49/4/0/6.
2026-08-08	Usage-scenario catalog delivered (`work/usage-scenarios/`, sessions 1–2, user brief): external scenarios E1–E18 (lifecycle, personal layer, working loop, knowledge access, concurrency, team capabilities) + internal scenarios I1–I10 (mechanism groups A–G with per-scenario test status), the zoom-level doc model (Z0–Z3 × O0–O4 + agent profiles), a scenario→docs→tests coverage matrix, and `gaps-and-coverage.md` (8 ranked gaps with recommendations + sequencing). Opened M15 (catalog meta-card) and promoted the three un-carded change-log findings to cards: M16 (EnterWorktree stales registry artifact — reproduced live this session), L32 (WARN push swallowed via busy child), L33 (launcher staleness race — shape recurred this session via a worktree branched from lagging origin/main). Scorecard now 4/49/4/0/6.
2026-08-08	Usage-scenarios session-4 walk-through + first two gaps shipped. User verdicts recorded on M15 (sequencing adopted as-is; `usage-scenarios.html` retirement endorsed; Gaps 6+2 green-lit; 3+8 and 7 queued; 4 and 1 deferred until real). Gap 6: new `scripts/tests/test-template-instantiation.sh` (20 asserts, nine suites now) — clean-room clone→`setup.sh`→`check-workspace-structure.sh` with produced-structure, idempotency, symlink-repair, and placeholder `--clone-repos` cases, plus doc-accuracy pins for the two claims previously caught false (`workspace-structure.md` registry↔disk reconciliation claim removed; `--clone-repos` ignores-tiers behavior now stated honestly). Gap 2: Z0 product interview checklist added beside the S0 preconditions in `workspace-structure.md`; `SPEC.md` and `docs/system-design.md` upgraded from bare stubs to fillable Z0 templates (components/entry-points tables); E1 local-vs-remote fork now an explicit choice point in `template-usage.md` with both paths. M15 stays Open (Gaps 3+8, 7 next). No scorecard change.
2026-08-08	Usage-scenarios session-5: Gaps 3+8 shipped (M15). Gap 3 (tooling manifest): `check-dependencies.sh` gains `req jq` (hard-required by `context-budget.sh`, previously unlisted) and a required hook-wiring check (I7's missing installation check — `.claude/settings.json` must wire the context-budget hooks); `check-service-access.sh` gains the same hard-fail distinction (required services → exit 1; was always exit 0); `recommended-tooling.md`'s blanket "everything optional" replaced with a "Required for everyone" table; runbooks updated (jq install lines, exit-code meaning); `setup.sh` dependency check moved after the per-user copies so its verdict reflects post-setup state. Gap 8* (capability authoring, documentation only): "Authoring a Team Capability" section in `workspace-structure.md` — container decision rule (`scripts/` vs `skills/` vs `.claude/agents/` vs `docs/runbooks/`), the two mechanical skill-wiring steps (commands mirror + `CONTEXT.md` listing), and the make-it-required hook into Gap 3's manifest; pointer added in `CONTEXT.md` → "Workspace Skills". M15 stays Open (Gap 7 last). No scorecard change.
2026-08-08	Usage-scenarios session-5 continued: Gaps 7+4+1 shipped, M15 closed. Gap 7: zoom model promoted from the work item to committed `docs/zoom-model.md`; superseded `usage-scenarios.html` retired to `docs/archive/` with a pointer stub; `README.md` repointed. Gap 4 (interface only, per the simplicity guardrail — document, don't build): team-vault interface + per-service `Scope` field documented in `docs/service-access.md`; no vault tooling. Gap 1 (phase 1 only): "Before You Add a Teammate" breakage points documented in `workspace-structure.md` + `user@host` identity fields in coordination records; phases 2–3 stay deferred by design until a second person is real (user verdict — not an open item). Also added a new `docs/README.md` index. All M15 gap work done → card Resolved and moved to the archive. Scorecard now 3/50/4/0/6.
2026-08-10	Fix M17 (new card, resolved on filing — archived): Copilot folder-trust seeding — closes the `work/usage-scenarios/` gap "trustedFolders step is checked by NO check script". `scripts/setup.sh` step 3b seeds `~/.copilot/config.json` `trustedFolders` idempotently (JSONC-aware: `//` comment lines stripped before `jq`, preserved on rewrite; already-trusted fast path; skip-with-note until Copilot's first run creates `~/.copilot/`; warn-not-fail on missing jq / unparseable config); `scripts/check-dependencies.sh` gains the verify-only advisory (trusts / does-NOT-trust with the silent-no-op warning); `docs/context-budget.md` vendor-table friction cell updated. `test-template-instantiation.sh` now sandboxes `$HOME` around `setup.sh` runs and adds a T7 seed/comment-preservation/idempotency block. Ported from the downstream insight workspace. Scorecard now 3/51/4/0/6.
2026-08-11	Fix: Copilot VS Code hook auto-registration (`6a6f1fa`, ported from the downstream insight workspace): the SessionStart/Stop hook script now registers before its `[ -f "$cs" ]` measurement guard — SessionStart fires seconds before VS Code creates `chatSessions/<sid>.jsonl`, so guard-first skipped registration on every fresh session and produced false "hook is a dead end" verdicts downstream; `register` handles the missing artifact as `method=deferred`. Live-verified on a fresh Copilot VS Code session (auto-registered in 1s, zero manual steps; VS Code 1.132.0, `.github/hooks/` confirmed a default `chat.hookFilesLocations` location). Also ported: Guard C in `context-budget.sh` (`register` dies rather than pinning a sibling when several Copilot sessions are active and `$VSCODE_TARGET_SESSION_LOG` is unset), `chat.useHooks` in `.vscode/settings.json`, docs/CONTEXT.md updates (auto-register primary, manual prefixed register demoted to hook-less fallback), and a ledger-evidence tightening of the `session-rollover` trigger policy (WARN = one small closing unit; pre-flight headroom check before major units) backed by the imported heavy-deployment analysis `work/context-decay/ledger-analysis-heavy-deployment-2026-08-11.md`. No scorecard change.
2026-08-11	Fix M18 (new card, resolved on filing — archived): rollover-cost optimization, from the cross-deployment question "the rollover itself costs ~20K — why, and how do we cut it". Analysis (`work/context-decay/rollover-cost-analysis-2026-08-11.md`): local cost tripled July→Aug (median ~4.6K → ~11K over 36 ledger deltas) tracking `session-rollover/SKILL.md` growth 4.6KB→12.1KB; floor ~8–10K mechanical + 0–25K deferred bookkeeping. Applied R1–R5: new `scripts/rollover-prep.sh` one-shot prep (start record, git summary, verified-lossless handoff→archive rotation, top block, `.session-seq`, options capture; 28-assert test suite), skill slimmed 12,074→9,222 bytes (mechanics disclosed to scripts), write-ahead promoted to standing discipline, ≤40-line handoff-block cap, self-sufficiency note (no `context-budget.md` load mid-rollover). First live run also caught up 4 unarchived blocks in `work/context-decay/handoff.md`. Scorecard now 3/52/4/0/6.
2026-08-11	Fix M19 (devex-review fix package (a) — archived): clean day-1 state. Telemetry ledger moved `work/context-decay/context-ledger.jsonl` → `.context-budget/context-ledger.jsonl` with a self-cleaning migration shim (any legacy ledger is folded in at script load; G3a–c regression asserts). Prune-`work/` guidance added as `template-usage.md` §5 (+ step-4 bullet in `setup-guide.html`): `work/` is a worked example, not machinery. Doc-worthy research promoted to `docs/archive/` (`ledger-analysis.md`, `ledger-analysis-heavy-deployment-2026-08-11.md`, `rollover-cost-analysis-2026-08-11.md`, `design.html` → `context-budget-design.html`) and every shipped-artifact citation repointed (`session-rollover/SKILL.md`, `context-budget.sh`, copilot-vscode hook, `context-inspect.sh`, `docs/context-budget.md` — whose pointer to the never-imported `copilot-vscode-hook-research-findings.md` was dropped as already dangling). All 10 suites green (402 asserts). Scorecard now 9 open / 54 resolved.
2026-08-11	Repair: L34 ID collision — the devex-review session (unpushed) and the downstream insight clone independently assigned L34 to different findings. The clone's card (onboarding canary + entrypoint/parameterization suites, filed resolved, `28c49f0`) published first and keeps the ID; the devex review's "context-budget minimal mode" card is renumbered L34 → L35 (references in the review header and `work/devex-review/decisions.md` updated; session-1/2 handoff history keeps the old ID and is explained by this row). Scorecard unchanged. Lesson for concurrent clones: an ID is only claimed once pushed.
2026-08-12	Fix M25 + L33 + L32 (devex-review fix package (d) — archived): session-mechanics reliability. M25: `rollover-prep.sh` splits root resolution — tracked work files (handoff rotation, git summary) now target the invoking checkout (`$PWD`'s toplevel, adopted only when it shares this repo's git common dir), untracked coordination state (`.session-seq`, `.rollover-options`) stays on the main checkout; sibling audit confirmed `capture-rollover-options.sh`/`context-budget.sh`/`launch-next-session.sh` touch only coordination state, so their main-checkout convergence is correct by design; worktree regression test T10. L33: mechanical stale-launcher guard in `launch-next-session.sh` — refuses to launch when any local/remote ref carries a `next-session.md`-touching commit not reachable from HEAD (reachability, not timestamps); `--skip-freshness` overrides; tests F1–F4. L32: claude hook keys throttle/escalation state per (session, chain) via transcript basename, so a sidechain firing can't swallow the parent's WARN; test T12. Suites green: 38+94+54 asserts. Scorecard now 3/61/4/0/6.
2026-08-12	Fix M23 + L35 (devex-review fix package (e), the program's final package — archived): non-engineer surfaces. M23: new `docs/for-non-engineers.md` (status / decisions / contributing routes + the six-term glossary, agent-prompt fallbacks for the optional to-spec/to-tickets toolchain); new `work/README.md` status index (one row + status per work item, refreshed at checkpoint/rollover) with a 60-second work-directory reading key; (anyone) audience tags on create-work-item/decision-log/to-spec in CONTEXT.md's skills list plus a human-redirect line atop CONTEXT.md; discoverability rows in `docs/README.md` (top-of-table non-engineer row; rendered backlog URL on the findings row) and a README.md pointer. L35: "Minimal mode" statement at the top of `docs/context-budget.md` (solo loop = register + record + session-rollover) with the fleet-only sections tagged in the section index. Doc-only change; no suites affected. Scorecard now 1/63/4/0/6 — remaining open card: M16.
2026-08-12	New findings M26 + L36 + L37 filed from an evidence pass over a real non-code instantiation (personal house-sale workspace, 8 sessions, no product code): the process core (launcher/ledger, decisions.md, measured rollovers) worked unmodified, surfacing three gaps — no-git privacy mode is real but unsupported (M26), instantiation pruning leaves template-development artifacts behind (L36), and the hand-invented "portable agent brief" export pattern has no convention (L37). Scorecard now 4/63/4/0/6.
2026-08-12	Fix M26 + L36 (house-sale mission, cards 1–2 — archived): M26: `docs/template-usage.md` §1 gains a third, no-git ("private") instantiation variant + a "No-git mode" subsection — paste-in Privacy Posture block for `CONTEXT.md`, Tier-1→Tier-2 decision-capture rewrite, N/A list for git-dependent machinery — with one-line N/A notes at each point of use (CONTEXT.md Tier-1 bullet; decision-log / checkpoint / session-rollover skills); no code change (`launch-next-session.sh` already degrades without git). L36: §5 broadened to the full template-development prune list (backlog pair, template-usage itself, unused mcp-fragments, LICENSE swap, README index rows, operational-knowledge trim; scaffolding dirs stay), `setup.sh` prune reminder gated on the backlog pair (regression T1c), M19 ledger-migration shim verified via registry G3. Suites: 72/72 registry, 28/28 instantiation. Scorecard now 2/65/4/0/6. (Row added retroactively by session 5 — the delivering session omitted it.)
2026-08-12	Fix L37 (house-sale mission, final card — archived): portable-agent-brief convention documented in `docs/work-directory-conventions.md` — new "Portable agent brief" section (one sealed file per version under `work/<proj>/briefs/<audience>-vN.md`, supersession header with the discard line, explicit never-reveal section, ledger notes at issue and at staleness, skeleton included) plus an optional-files-table row; contrasted with dispatch records (same-machine children). Kept convention-only per simplicity-first — a skill is promotable if the pattern recurs (Tier-2 note). Scorecard now 1/66/4/0/6.
2026-08-13	New findings M27 + M28 + M29 + L38 filed from the SDLC map's gap register (`work/sdlc-ai-mapping/sdlc-map.md`, dispositions agreed 2026-08-13): design-time testability prompt (G4→M27), UAT/beta coordination (G5→M28), postmortem convention (G6→M29), dependency/test-suite health (G8→L38). The two High-severity gaps became work items instead (`work/feedback-intake/` for G1, `work/quality-gates/` for G2+G3); G7 (release engineering) was scoped out deliberately. Scorecard now 5/66/4/0/6.
2026-08-13	New finding L39 filed from the SDLC map's round-2 review fix pass (finding F10, `work/sdlc-ai-mapping/review2-findings.md`): the map's N8 "backlog conventions (native)" claim covered an unshipped capability — the template repo's own backlog practice is template-repo-only and deleted on adoption. Map rescoped to the issue-tracker conventions that do ship; the generic-backlog-convention gap is now G9 in the map's gap register → this card. Scorecard now 6/66/4/0/6.
2026-08-14	Fix L40 (filed resolved — archived; ported from the downstream Insight workspace's context-efficiency audit): fragment bring-in mechanics added to `mcp-fragments/README.md` and `docs/mcp-setup.md` (subagent `mcpServers` frontmatter, `ROLLOVER_OPT_EXTRA` fragment hand-off, no-attach-while-running + schema-deferral facts, variadic `--mcp-config` prompt-ordering trap); `disableClaudeAiConnectors: true` added to `.claude/settings.json.example`. Filed downstream as L35; renumbered at publish because L35 was already claimed (context-budget minimal mode) — second instance of the L34 lesson: an ID is only claimed once pushed. Scorecard now 6/67/4/0/6.
2026-08-16	New finding M30 filed and resolved in the same pass: the Matt Pocock skill set was global-install-only, so template downloaders got 5 of 27 skills. Full curated set now vendored at upstream `068b6e0` with `scripts/sync-vendored-skills.sh`, `skills/vendored-skills.md` index (+ MIT license), 10 slash wrappers, and a rewritten `recommended-tooling.md` §3 (vendored = default, global symlinks = maintainer path). Scorecard now 6/68/4/0/6.
2026-09-09	Fix M38 (downstream contribution, PR #44 `fix/session-loop-notify-hook` — filed resolved, archived; this row added by the pulling session per the maintenance convention): supervised-chain halts were easy to cause and miss. Red-flag block after step 7 in `skills/session-rollover/SKILL.md` (the "rollover complete" record is not the sentinel); stock `scripts/session-loop-notify.sh` notify hook (desktop notification + echo fallback) with suite `test-session-loop-notify.sh`; `docs/context-budget.md` documents the `SESSION_LOOP_NOTIFY` wiring and the load-bearing `${ROOT:-.}` guard. Scorecard 5/81/4/0/6 at the time.
2026-09-10	Fix M27 + M28 + M29 + L38 + L39 (the SDLC-map design-gap cards — archived; `dc1f334`, `work/template-improvement-review/` session 1, built under stated assumptions after the user authorised autonomous improvement; the walk-through of `work/template-maintenance/open-cards-options-brief.md` is thereby settled). M27: native advisory skill `design-for-testability` + optional spec `## Testability` heading. M28: `uat.md` slot in the work-directory conventions. M29: committed `docs/postmortems/` (blameless README + template, feedback-intake forward pointer). L38: routed into `work/quality-gates/` scope. L39: "bring your own tracker" posture in `docs/agents/issue-tracker.md`, naming these backlog files as a copyable example. Scorecard now 0/86/4/0/6.
2026-09-10	New finding L45 filed from the pooled review (`work/template-improvement-review/review.md` A10): gitignored work directories are absent from the repo guard's forced worktree and lost when it is deleted, so background sessions on local-only items need a manual copy-back. Scorecard now 1/86/4/0/6.
2026-09-10	Fix L45 (archived; `work/template-improvement-review/` session 3, branch `fix/l45-gitignored-work-dirs`): local-only (ignored) `work/<item>/` dirs are now shared into git worktrees by symlink — new `scripts/link-local-work.sh`, called unthrottled from the shared hook lib (every runtime's per-tool hook) and from `context-budget.sh register`; registers the exact path in `.git/info/exclude` so the symlink never shows as untracked. The card's "repo guard" turned out to be the runtime's own worktree isolation, so exemption was impossible and copy-back was lossy (auto-cleaned "unchanged" worktrees) — decision in the item's `decisions.md`. Suite `test-link-local-work.sh` (30 asserts); docs in `work-directory-conventions.md` + `operational-knowledge.md`. Scorecard now 0 open / 87 resolved. Session 4: probed live with `Agent(isolation: worktree)` (link, write-through, exclude line, survival past auto-clean all confirmed; result noted in `operational-knowledge.md`) and merged to main `5fd5480`.
2026-09-13	Integrated origin PRs #45–#53 into this instance (merge commit; rows added here because the delivering commits carried no backlog update): #45 `diff-review.sh -w/--worktree` reviews the uncommitted working tree; #46 launcher convention — mandatory work must sit in the First-actions block (conventions doc + `session-rollover`); #47 session-loop round 2 — bump record is the rollover verdict (`rollover-complete` sentinel retired from docs), vendor-logout classification, persisted chain budget with `--reset-cap`; #48 duplicated notify-hook block dropped from `context-budget.md`; #49 TADR-0009 indexed; #50 `create-work-item` registers the new project in `work/README.md`; #51 `session-rollover` says to update specialized state files; #52 `checkpoint` gains untracked-`work/` check, launcher re-read, ADR scope rule, handoff/decision-log as prerequisites; #53 Copilot entrypoint `.github/copilot-instructions.md` shipped (entrypoint registries take `path:target` pairs). Merge conflicts: L45's `link-local-work` call kept alongside `resolve_runtime_pid` in `cmd_register`; scripts tree kept both new lines. No card-count change.
2026-09-16	Template changes pulled from `work/template-improvement-review/` Stage 4 phase 0 (session 10), no card: root `ROLLOVER_RELAUNCH` default flipped `auto` → `manual` while the session scripts are rebuilt (per-item override keeps that item on `auto`); `SESSION_LOOP_NOTIFY` now resolves from the env file's own location via `BASH_SOURCE` instead of the `${ROOT:-.}` guard, which resolved against the cwd for every caller but the supervisor (`test-session-loop-notify.sh` N4); `jq` pinned as a hard requirement (`test-check-dependencies.sh` D5); one-time `scripts/import-session-seq.sh` + suite, to be retired after cutover. Scorecard unchanged.
2026-09-14	New finding L46, fixed in the same pass (archived; `work/template-improvement-review/` session 5): the session-loop round-2 runtime files `.session-seq.bump.json`, `.session-loop.budget` and `.session-loop.alarm-stop` were never added to `.gitignore` (commit `a91c86c` / PR #47 added the writers only; `test-session-loop.sh` models an ignore list the template did not ship), so every supervised chain left a dirty tree and a `git add -A` session would defeat the progress guard. Three patterns added. Scorecard now 0/88/4/0/6.
2026-09-22	New finding L47, fixed in the same pass (archived; `work/template-improvement-review/` session 22): the counter-import script was slated for retirement after the Stage 4 cutover, but every downstream workspace pulling the new scripts has work items on the old `.session-seq` counter — kept it, added `--status [<project>]` (fresh / old / imported / new / conflict, which generation of `session-loop.sh` ran, leftover files), tests, and a migration table in `docs/context-budget.md`; the six remaining live items imported. Opened M39 (attended `register` after a stop-door `close` re-binds the closed session's number). Scorecard now 1/89/4/0/6.
2026-09-22	Resolved M39 (archived): `bind_record` now mints `seq + 1` when the dead owner left through the stop door (outcome `minted`, logged with the loser), matching the launcher's `stopped` path; plain-exit adoption unchanged. Regression R6i–k; both docs say it in a sentence. Scorecard now 0/90/4/0/6.
2026-09-22	New finding M40, fixed in the same pass (archived; `work/session-management-followups/` ticket 01): `attach-session.sh` and the statusline resolved an item's owner through the retired `.active-session` lock, so every item read as unowned. Both now read the record's `session` block through two new `session-lib.sh` readers with the TADR-0010 pid liveness rule; suites rewritten onto record fixtures.
2026-09-22	New finding M41, fixed in the same pass (archived; `work/session-management-followups/` ticket 02): the ledger checker and the launcher's `top_ledger_session` disagreed on which headings carry a session number (the numbered addendum form passed the gate, then refused `ledger_shape` at rollover). One rule now stated in `docs/work-directory-conventions.md` → Ledger, both parsers cite it, and `test-check-ledger.py` runs a heading fixture set through both. Scorecard now 0/92/4/0/6.
2026-09-22	New finding M42, fixed in the same pass (archived; `work/session-management-followups/` ticket 03): no workspace-level template version marker existed. Root `TEMPLATE_VERSION` (one ISO date behind `#` comments, bumped at each template release, read by nothing yet), named in `docs/template-usage.md` §6 "Upgrading later" and the structure tree; `test-template-version.sh` proves it parses and survives the §5 prune. Shape recorded as an assumption in the item's `decisions.md`. Scorecard now 0/93/4/0/6.
2026-09-23	Added L48 (wayfinder as a plan template) from the `work/plans` build verdict; deferred until the first real plan has closed.
2026-09-30	Template change from `work/plans/` ticket 12, no card (dogfood finding s16): `scripts/check-ledger.py` now fails on a block heading filed twice (a redone ledger write that left its stale first pass); `test-check-ledger.py` gains fixtures for both s16 shapes (block spliced into the purpose comment — already caught — and the duplicate) plus a legal same-day pair in the clean baseline; `skills/checkpoint/SKILL.md` step 2 now anchors ledger writes on `-->` and runs the checker, as `session-rollover` already did.
2026-09-30	Template changes from `work/plans/` tickets 13–15 (dogfood findings), no card: a closed `chain.plan` yields to the one open plan (`plan.sh`, `session-loop.sh`); `plan.sh` ergonomics — `add --leaf` implies `tier: auto`, `--project=`/`--plan=` forms, a `relative-check` lint rule, `done` shows check output only on failure; `docs/plans.md` and `skills/plans/SKILL.md` wording (check cwd, sizing, closing a plan, `sync` after `done`, hitl launcher prose) and the L48 recipe “Create a plan before a spec exists”. `TEMPLATE_VERSION` bumped to 2026-09-30. Resolved L48 (archived): adopt as a recipe, keep the wayfinder skill. Scorecard now 0/94/4/0/6.
2026-09-24	Feature: plans — a work item's work as a node graph (`work/<item>/plans/NN-<slug>/`, one markdown file per node, waves closed by a reconcile node, `hitl` nodes for a person's step), driven by `scripts/plan.sh` (status / frontier / check / start / done / sync and the rest; every state write) and the vendor-neutral `plans` skill (`/plan` on Claude Code). Tier policy and per-runtime model knobs in `plan-tiers.env` (no model name in a plan file); the supervisor (`session-loop.sh`) runs `frontier`/`check` between children. Landed over tickets 01–10 of `work/plans` (feature work tracked in the work item, recorded here per the delivering-commit rule). Front door: `CONTEXT.md` → Plans + the plan glossary under Language (worktree as an alias to avoid); reference `docs/plans.md` (format, verbs, tiers, replans, wayfinder mapping, per-runtime table, worked example), indexed from `docs/README.md`; optional `plans/` row in `work-directory-conventions`; `for-non-engineers` glossary entry. Suites: `test-plan.sh`, `test-session-loop.sh`, and `test-doc-consistency.sh` now pins every path the plans docs name. `TEMPLATE_VERSION` bumped to 2026-09-24. L48 (wayfinder as a plan template) stays open, gated on the first real plan closing. No scorecard change.
2026-10-07	New findings M43, L49, L50, L51, D5 from the context/memory-engineering evaluation (`work/context-memory-eval/eval.md`, recommendations 1–4 and 6 accepted; 5 deferred). Work tracked in `work/context-memory-hardening/`. Scorecard now 5/94/4/0/6.
2026-10-07	Resolved L49 (archived; `work/context-memory-hardening/` ticket 01): “Cache the prefix, vary the tail” rule in `docs/context-budget.md` with a `CONTEXT.md` pointer, a date-line warning in `check-workspace-structure.sh` (new `test-context-prefix-stability.sh`), and a per-runtime hook-output audit (no change needed). Scorecard now 4/95/4/0/6.
2026-10-07	Resolved L50 (archived; `work/context-memory-hardening/` ticket 02): ledger rows carry `cache_read_share` from the runtime's own cache count (Claude, Codex, OpenCode; `null` elsewhere), shown as `cache=NN%` in `check`/`record`. Scorecard now 3/96/4/0/6.
2026-10-07	Resolved M43 (archived; `work/context-memory-hardening/` ticket 03): `docs/operational-knowledge.md` entries carry git-derived `Last confirmed` dates with a 6-month review age; `checkpoint` and `decision-log` classify a new entry ADD / UPDATE / SUPERSEDE / NOOP before writing, and `checkpoint` lists stale entries for a person. Scorecard now 2/97/4/0/6.
2026-10-07	Resolved L51 (archived; `work/context-memory-hardening/` ticket 04): `CONTEXT.md` trimmed from 20,656 to 15,206 bytes by moving detail behind pointers to the docs that already hold it; every must-see rule stays inline. Scorecard now 1/98/4/0/6.
2026-10-07	Decided D5 (archived; `work/context-memory-hardening/` ticket 05): TADR-0013 — no vector store, no MMR re-ranking, no numeric decay scoring; M43's last-confirmed dates are the substitute. All five context/memory-hardening cards closed. Scorecard now 0/98/5/0/6.
2026-10-07	New finding L52: `skills/session-rollover/SKILL.md` cites a `CONTEXT.md` section (“Recording new learnings”) that doesn't exist. Scorecard now 1/98/5/0/6.
2026-10-07	Resolved L52 (archived): dropped the dead `CONTEXT.md` → “Recording new learnings” pointer in `skills/session-rollover/SKILL.md` and repointed two mentions of the nonexistent `scripts/check-tooling.sh` to `scripts/check-dependencies.sh`. Scorecard now 0/99/5/0/6.
2026-10-08	New findings L53–L59 from the doc gap review (`work/context-memory-hardening/doc-gap-review.md`), and M44, M45, L60–L62 from a harness-engineering review of Böckeler, “Harness engineering for coding agent users” (martinfowler.com); M44–M45 and L60–L62 are built in `work/harness-engineering/`, L53–L59 in `work/template-maintenance/`. Scorecard now 12/99/5/0/6.
2026-10-08	Resolved L53 (archived): `docs/context-budget.md` check/record output line gains `cache=`. Scorecard now 11/100/5/0/6.
2026-10-08	Resolved L55 (archived): `docs/workspace-structure.md` → “What `CONTEXT.md` Should Contain” now states the Z0 size budget and the no-dates rule. Scorecard now 10/101/5/0/6.
2026-10-08	Resolved L56 (archived): two stale `CONTEXT.md` → “Tool & Context Loading” pointers fixed in `docs/workspace-structure.md` and `docs/mcp-setup.md`. Scorecard now 9/102/5/0/6.
2026-10-08	Resolved L57 (archived): retired decision notes explained in `docs/work-directory-conventions.md` and `docs/for-non-engineers.md`. Scorecard now 8/103/5/0/6.
2026-10-08	Resolved L58 (archived): the ledger-splice gotcha in `docs/operational-knowledge.md` no longer sends readers to the removed `scripts/rollover-prep.sh`. Scorecard now 7/104/5/0/6.
2026-10-08	Resolved L59 (archived): TADR-0009 amendment repointed from the never-committed `docs/session-chain-scenarios.md` to the `relaunch_off` gate and its test. Scorecard now 6/105/5/0/6.
2026-10-08	Resolved M44 (`work/harness-engineering` ticket 01): `scripts/run-checks.sh` (full + `--fast` tiers), opt-in pre-commit hook, GitHub Actions workflow (unverified until pushed); `test-jev.sh` made hermetic. Scorecard now 5/106/5/0/6.
2026-10-08	Resolved M45 (`work/harness-engineering` ticket 02): `scripts/check-drift.sh` staleness sweep, run by checkpoint step 1 and by CI on push and weekly. New L63 (open): two removed scripts still listed in the `workspace-structure.md` tree. Scorecard now 5/107/5/0/6.
2026-10-08	Resolved L60 (`work/harness-engineering` ticket 03): checkpoint gains “Promote a repeat to a check”; gotchas may carry an `Enforced by:` line, back-filled on three entries. Scorecard now 4/108/5/0/6.
2026-10-08	Resolved L62 (`work/harness-engineering` ticket 05): every new check ships with a suite that makes it fail; three new suites close the gaps the audit found. Scorecard now 3/109/5/0/6.
2026-10-08	Resolved L61 (`work/harness-engineering` ticket 04): one table of the workspace's guides and checks in `docs/workspace-structure.md`; its gaps filed as L64 and L65. Scorecard now 4/110/5/0/6.
2026-10-08	Resolved L63: two removed scripts dropped from the `workspace-structure.md` tree; `check-workspace-structure.sh` now checks the tree against disk. Scorecard now 3/111/5/0/6.
2026-10-08	Resolved L64: `scripts/check-plans.sh` puts `plan.sh check` for every committed plan in the gate (fast tier). Scorecard now 2/112/5/0/6.
2026-10-08	Resolved L65: `test-agent-entrypoints.sh` case E5 checks every `SKILL.md` has `name`/`description` frontmatter. Scorecard now 1/113/5/0/6.
2026-10-08	Resolved L54: `TEMPLATE_VERSION` bumped to 2026-10-08 for the release push. Scorecard now 0/114/5/0/6.
2026-10-09	Refreshed the vendored Matt Pocock skills from `068b6e0` to `49dd158`. Added `implement-spec`, `pr` and `retro`. Dropped `resolving-merge-conflicts`, which upstream removed. Moved the glossary from `CONTEXT.md` into a new `GLOSSARY.md`, following upstream's rename (`work/template-maintenance/decisions.md`).
2026-10-09	Filed M48 (`isolated: yes` has no effect) and L71 (plans for code work have no review step), both from the implement-spec vs plans comparison. Scorecard now 3/120/5/0/6.
2026-10-09	Resolved L70: new patched class of vendored skill (upstream copy plus a `workspace.patch` the sync re-applies, failing loudly when it no longer applies), with `scripts/tests/test-sync-vendored-skills.sh`. `code-review` is the first: its patch groups each axis into blockers and suggestions. Scorecard now 2/121/5/0/6.
2026-10-09	Resolved M48 (what an orchestrator does with `isolated: yes`: own worktree and branch, merged back before `done`) and L71 (a plan that changes code ends with a `code-review` wave), both in `skills/plans/SKILL.md`. Scorecard now 0/123/5/0/6.
2026-10-09	Resolved L72 (TADR-0014: the glossary lives in `GLOSSARY.md`), M49 (`scripts/migrate-glossary.sh` + runbook for older projects) and L73 (patches name their upstream issue; the sync flags closed ones; `check-drift` warns on vendored skills over 60 days old). Opened L74 (intermittent `test-session-lib.sh` S8 failure). Scorecard now 1/126/5/0/6.
2026-10-09	Resolved L74: a second lost race in the session-record lock (a lock re-taken between two checks read as `record_unwritable`); deterministic test S10r–t. Scorecard now 0/127/5/0/6.
2026-10-11	Session 24: analysed Addy Osmani's `agent-skills` and pstack against the vendored Matt Pocock set (`work/template-maintenance/skill-packs-analysis.md`): take 16 of 77 skills, skip the rest as duplicates, Cursor-bound or domain-specific. Opened M50 (multi-upstream sync + drift) and L75–L81 (vendoring in three waves, catalog + wiring, round-trip proof). Scorecard now 8/127/5/0/6.
2026-10-11	Resolved M50 (one sync script and drift check for three upstreams), L75–L79 (16 skills vendored: 9 Addy Osmani, 7 pstack; patches for the runtime-specific lines), L80 (catalog, `CONTEXT.md` pointer, docs, slash commands) and L81 (sync round trip clean, audit zero blockers, 42 checks pass). Scorecard now 0/135/5/0/6.
2026-10-11	Session 24, second task (Rob Shocks' “Pocock vs Pstack vs Osmani”): opened M51 (trigger trimming + description budget check), L82 (stakes-to-ceremony guide) and L83 (anti-rationalization tables); plan `work/template-maintenance/plans/01-skill-ceremony`. Scorecard now 3/135/5/0/6.
2026-10-11	Resolved M51 (unslop narrowed, 32 model-invoked skills reviewed, three made user-invoked, `check-skill-descriptions.sh` ratchets at 7,500 bytes; 10,174 → 7,303), L82 (`docs/stakes-to-ceremony.md` + pointers) and L83 (rationalization tables in five workspace skills). 44 checks pass; sync round trip clean. Scorecard now 0/138/5/0/6.
2026-10-12	Resolved L84 (from the token-factory sync of `ccb4ed3`): template ADRs move to `docs/template-adr/` as `TADR-NNNN`, `docs/adr/` is the project’s own; instances propose TADRs upstream via `Promote?: upstream`; migration runbook `docs/runbooks/adr-namespace-migration.md`. Scorecard now 0/139/5/0/6.
2026-10-12	Test only: `test-session-loop.sh` gains W1 — the supervisor run from a worktree’s own copy of `session-loop.sh` drives the main checkout’s record and writes nothing into the worktree (the `--git-common-dir` resolver had no loop case from a worktree; seen red with the resolver forced to `SCRIPT_ROOT`). Replaces PR #61, whose other three worktree cases tested the `.next-command` pair and `--emit <path>` that Stage 4 removed. No card-count change.
2026-10-12	Test only: `test-attach-session.sh` stubs `claude` on PATH (`agents --json` → `[]`); on a machine running a live session named `testproj #N`, T5a/T5b resolved it and exited 0 (seen red 25/2, now 27/0). Opened L85 (template-only suites should skip in instances). Scorecard now 1/139/5/0/6.
2026-10-12	Added L86 (Open): Antigravity IDE/app and Codex desktop app session-identity probes deferred from runtime-session-identity node `11-remaining-surface-probes` on the user’s call.
2026-10-12	Added L87 (Open): cwd substring match in context-budget discovery, from the runtime-session-identity node 15 review. The same review fixed three items in `scripts/context-budget.sh` and `.claude/settings.json`, test-first (commits `0b4dd7c`, `0d581ce`, `0964a23`).
2026-10-12	Resolved L88 (PR #65, pulled from origin without a card: stray files in `nodes/` no longer break `plan.sh`) and L89 (`plan.sh plans` lists every work item’s plans, one line each, `--json`). Scorecard now 3/141/5/0/6.
2026-10-12	Resolved L87 (cwd substring match in context-budget discovery): both checks now match the cwd as a whole value; registry suite U3/U4 seen failing first. Card moved to the archive.
2026-10-18	Resolved L94 (external-tools-review doc pass, R7–R10, R13–R17, R20, R22–R24) and L93 with it (yt-dlp example reworded, R6). Cards moved to the archive. Scorecard now 6/147/5/0/6.
2026-10-18	Resolved L85 (template-only suites skip in instances), L95 (dispatch-contract read-first / Remember: / WIP commits / proved DONE) and L96 (usage-limit wait vs the silence watchdog, documented) — backlog-sweep wave 1. Cards moved to the archive. Scorecard now 3/150/5/0/6.
2026-10-18	Resolved L98 (skill-earns-its-place pilot: 3 skills, 36 headless runs, ~$0.88; no skill met the keep rule). Opened L99 (`code-simplification`) and L100 (`unslop`) as prune candidates for the user to decide. Scorecard now 4/151/5/0/6.
2026-10-18	Fix (no card; runtime-session-identity plan 02, closed): Copilot VS Code agent mode on the Copilot CLI runtime context budget. The exact count comes from the CLI checkpoint's `prompt_tokens`, and the CLI cumulative overcount is gone (`3928a86`). Claude `SessionStart`/`SessionEnd` hooks skip Copilot session-state transcripts (`34c76b4`). There is one registry name per Copilot session, copilot-cli (`6c5ae3a`), and the name is keyed on the transcript path, so a leaked env var cannot re-file a VS Code Local chat (`f87afb0`, from the plan’s code review). Each fix has a suite case seen red first. Live re-probe in VS Code passed.
2026-10-18	Resolved L86 (Antigravity IDE and Codex desktop probes both pass; the Codex probe led to the Claude-hook allowlist, `cc1173e`). Opened L101 (manual `register` in the Antigravity IDE shell). Card moved to the archive. Scorecard now 5/153/5/0/6 (with L102/L103 from html-kit).
2026-10-18	Resolved L99: pruned `code-simplification` from the vendored set (user’s call). Card moved to the archive. Scorecard now 4/154/5/0/6.
2026-10-18	Resolved L100: kept `unslop` after a harder-prose re-test (37/37/52 of 60). Card moved to the archive. Scorecard now 3/155/5/0/6.
2026-10-18	Resolved L101: manual `register` in the Antigravity IDE shell now finds the transcript under any data root when `ANTIGRAVITY_APP_DATA_DIR` is unset; registry suite L2d seen red first. Card moved to the archive. Scorecard now 2/156/5/0/6.
2026-10-21	Resolved L103: `build-guide-html.sh` emits the HTML kit skeleton (no inline style; pandoc classes and `style=` stripped); `test-build-guide-html.sh` seen red first. Card moved to the archive. Scorecard now 1/157/5/0/6.
2026-10-22	Resolved L104: `session-loop.sh` no longer syncs a plan that is already closed, so a closing session’s launcher without position markers ends `plan_closed`, not `plan_invalid`; session-rollover keeps the markers after a plan closes. Loop suite P3h seen red first. Filed and resolved at once; card in the archive. Scorecard now 1/158/5/0/6.