orc-pack v1.13.1

[1.13.1] - 2026-10-06

Summary. Orc's paceline bar now has one cell per task plus a final ready cell, so a four-task run draws five cells and the active cell shows where the run is. Fix rounds are stages inside their task's cell (fix 1 to fix 3) rather than extra cells, the step label names what each task builds, the run is named after its issue (orc #42), and the ready check and landing have a cell of their own instead of a bare label. When a task is added mid-run, orc skips the old ready step and appends the new tasks and a fresh ready step, so the ready cell stays last.

Changed areas.

• skills/orc/SKILL.md, Progress bar section — new plan shape and update rules: per task stages ["build", "review", "fix 1", "fix 2", "fix 3"] (with design first on structural tasks), a descriptive label such as t2 upload limit, a final ready step with stages ["check", "land"], and the skip-and-re-add rule for new tasks. Refresh.
• README.md — Progress on your status line describes the new bar. Kit docs only.
• Every skill and agent file — pack: marker moved to orc-pack@1.13.1.

Update steps. Refresh skills/orc/SKILL.md and move every pack: marker to orc-pack@1.13.1. No new or removed files. An install coming from 1.12.x or earlier also applies [1.13.0] and every entry in between.

Breaking changes. None. A run started by 1.13.0 that is still on the status line keeps its old shape until the next orc run replaces it.

Files to read. skills/orc/SKILL.md (Progress bar).

---

orc-pack v1.13.0

[1.13.0] - 2026-10-06

Summary. Orc can now show a run's progress on the Claude Code status line through paceline, whose paceline-mcp server gives agents progress tools. After planning, orc starts a bar with one step per task, weighted by size and moved through each task's stages (design on structural tasks, then build, review, and commit), so the percentage tracks the real work. The integration is optional, and each update rides along with a call orc already makes, so it adds no turns.

Changed areas.

• skills/orc/SKILL.md, new Progress bar section (after The task list is your spine) — the calls orc makes at each point of a run, and when it skips the bar: the tools are missing, a call fails, or the invocation says the caller reports its own progress through paceline, which keeps one run per repository. A caller that draws its bar some other way needs no change; both bars show. Refresh.
• README.md — new Progress on your status line section linking to paceline. Kit docs only.
• Every skill and agent file — pack: marker moved to orc-pack@1.13.0.

Update steps. Refresh skills/orc/SKILL.md and move every pack: marker to orc-pack@1.13.0. No new or removed files in the install. Installing paceline is optional and outside the pack: put the paceline and paceline-mcp binaries on the machine, run paceline install, and register the server with claude mcp add --scope user paceline -- /path/to/paceline-mcp. A skill or loop that invokes /orc and reports its own progress through the paceline tools should say so in the invocation, or orc's bar replaces it. An install coming from 1.11.x or earlier also applies [1.12.0] and every entry in between.

Breaking changes. None. With paceline registered, orc runs now write <git-dir>/paceline/progress.json through its MCP server in the repo they work in.

Files to read. skills/orc/SKILL.md (Progress bar), and paceline's README, section "Progress bar", for the tool and file format.

---

orc-pack v1.12.0

[1.12.0] - 2026-10-06

Summary. Orc now uses Fable 5.1 where it still beats Opus 5.5: the ui-implementer runs on Fable for the design brief and first build of a structural UI task, such as a new screen or UI feature, because Fable's open-ended visual design is stronger. Everything else stays on Opus 5.5, which matches or beats Fable on coding at well under half the cost. Round three of a fix loop still runs on Fable, and a round-three fix that weakens a test is now a Blocker. Orc also records what every dispatch costs in a token ledger and reports the run's total, and it skips the verifier in every mode when a review returns only Nits. This release also trims repeated text across the pack, so each agent loads fewer tokens with no rule removed, and fixes several bugs: fallow could report a crash as a pass, dependency-vetter's audit failed without a lockfile, impact miscounted lines, and the suppression rules contradicted each other.

Changed areas.

• skills/orc/SKILL.md, Model selection and 5. Run the task loop — new Fable rule: ui-implementer gets model: fable for the brief and first build of a structural task; its fix rounds stay on Opus 5.5. Round three keeps Fable, and a round-three fix that weakens, skips, or deletes a test or assertion is a Blocker. In efficiency mode, structural UI stays on Opus 5.5. The per-agent tier list is replaced by one line naming the non-Opus agents, because each agent's frontmatter holds its model and effort. Refresh.
• skills/orc/SKILL.md, new Token ledger section and 9. Report — orc appends one tab-separated line per dispatch (date, run, task, size, agent, model, role, tokens, findings raised, findings confirmed) to .git/orc/token-ledger.tsv, found with git rev-parse --git-path, so it is never committed and persists across runs. The report gains a Cost line: the run's subagent tokens, dispatch count, three costliest agents, and the ledger path. Refresh.
• skills/orc/SKILL.md, 5. Run the task loop — the verifier is skipped in every mode, not only efficiency mode, when a panel returns no findings or only Nits with none tagged Cleanup candidate; those Nits go in the commit body marked unverified. A cleanup audit is always verified. The efficiency-mode question no longer lists this as a trade-off. Refresh.
• skills/orc/SKILL.md, other sections — shorter Waiting on subagents, Discoveries, and What orc does NOT do (rules stated elsewhere removed, no rule lost). Ready and the Standards table now allow a suppression only as code.md allows it. Refresh.
• skills/orc/references/builder-contract.md — step 10 defers to code.md's "Suppressions are a last resort" rule instead of banning every suppression; skipped tests, deleted assertions, and --no-verify stay banned. Shorter Standards list and step 8. Refresh; no pack: marker.
• agents/fallow.md — captures the exit code and stderr instead of discarding them; a new ERROR status covers any exit other than 0 or 1 or output that isn't JSON, so a crash is never a PASS. Defines PASS, WARN (dead code, duplication, complexity), and FAIL (a circular dependency). Anything that parses its report must accept ERROR and the new Exit code line. Refresh.
• agents/dependency-vetter.md — vets the exact version the caller names, resolving latest, another tag, or a range to a concrete version first. Creates the vet directory before entering it, and builds a lockfile (npm install --package-lock-only --ignore-scripts, then npm ci --ignore-scripts) so npm audit works. Refresh.
• agents/impact.md — counts with git diff -w --numstat (one diff against HEAD for uncommitted work, plus untracked files), so whitespace-only changes and binary files drop out and per-file counts are exact. Refresh.
• agents/security-reviewer.md — the output format gains the Pre-existing section its rules already asked for. Refresh.
• agents/test-coverage-reviewer.md — description now covers every change to logic that can regress. Refresh.
• skills/orc/references/standards/ui.md and agents/ui-reviewer.md — 768 px stays a design target; the reviewer renders 390 px and 1440 px and checks the tablet width by reading the code. Refresh.
• skills/newissue/SKILL.md — writes the issue body to a file and files it with --body-file, never a heredoc; the self-check now comes before filing. Refresh.
• skills/update-orc/SKILL.md — also finds a global install under ~/.claude/ when the project has none; fetches the kit before building the update map. Refresh.
• Token trims, no rule removed — agents/{implementer,ui-implementer,api-implementer,data-implementer,architecture-reviewer,design-reviewer,verifier,typecheck,next-issue-finder,docs-writer,skill-vetter,orc-updater}.md, skills/orc/references/{design-brief.md,standards/code.md,standards/comments.md,standards/data.md}, and every framework and platform playbook. Builders now point to the standards they load instead of restating them. design-brief.md gains a Reversibility bullet in its Data section. Refresh all.
• INSTALL.md, UPDATE.md, README.md, CLAUDE.md, docs/model-tiers.md (new) — install table uses one row for all agents and lists every kit file that isn't copied; stale statements fixed; the model-tier evidence moved from CLAUDE.md to docs/model-tiers.md. Kit docs only; don't copy docs/.
• Every skill and agent file — pack: marker moved to orc-pack@1.12.0.

Update steps. Refresh every skill, agent, and reference file listed above and move every pack: marker to orc-pack@1.12.0. No new or removed files in the install. Fable 5.1 must be available on the account for the structural UI move; on an account without it, those dispatches may fail, so remove the structural UI rule from Model selection there. If the install has local edits to fallow.md, keep the new exit-code capture. An install coming from 1.10.x or earlier also applies [1.11.0] and every entry in between.

Breaking changes. None removed. Orc runs now write .git/orc/token-ledger.tsv in the repo they work in, and the report gains a Cost line. Structural UI tasks now cost more per build (Fable 5.1 is priced above Opus 5.5). The fallow report can now say Status: ERROR.

Files to read. skills/orc/SKILL.md (Token ledger, Model selection, Efficiency mode, 5. Run the task loop, 7. Ready, 9. Report), skills/orc/references/builder-contract.md (Build mode step 10), agents/fallow.md, agents/dependency-vetter.md, agents/impact.md, and docs/model-tiers.md in the kit for the reasoning.

---

orc-pack v1.11.0

[1.11.0] - 2026-10-06

Summary. Every orc run now works to a written work contract, and a new scope-reviewer holds each diff to it. Before it plans, orc writes one sentence naming the exact files and symbols it believes the request means, what's in and out of scope, every integration marked real or mocked, its user-facing assumptions, and three to six acceptance criteria that each name a test. Builders write those tests first and show them failing before they write the change, and the report maps each criterion to its passing test. This targets the two costliest failure modes of unattended runs: a wrong approach (for example, a mock where the real API was wanted) and a misread request. /orc contract first … adds an approval stop before building. A new optional skill, /shipcheck, verifies a pushed commit all the way to the live build: it waits for CI and fixes a red run, checks that the target environment has every required secret, waits until the deployment serves that exact commit, and runs browser smoke checks at phone and desktop widths in both themes. Separately, undirected orc runs stop picking issues that are already fixed: the issue scout checks each candidate against the integration branch's history first, and orc closes the stale ones with evidence.

Changed areas.

• skills/shipcheck/SKILL.md (new) — run after a push. Watches the commit's CI runs with gh run watch; on a red run it reproduces the failing command locally, dispatches a builder to fix it, and re-pushes, at most 2 attempts, never onto main, master, or production. Compares the secret names in .claude/required-secrets.md with what the target environment has (GitHub Actions, GitHub environments, Vercel, Cloudflare Workers), never reading a value, and stops on any missing name. Proves the deployment serves the pushed SHA before testing. Runs every flow in .claude/smoke-checklist.md in Claude in Chrome at 390px and 1440px in light and dark mode, with a screenshot per step. Stops with exact unlock instructions at an SSO, auth, or VPN wall. Turns each confirmed regression into a fix/<slug> branch with a reproducing test (pushed, never merged), or an issue with screenshot paths and a root-cause hypothesis. Both .claude/ files are repo-owned; shipcheck drafts them from the code when they are missing. Copy it in for a copied install.
• skills/orc/references/work-contract.md (new) — the contract template: restatement, in scope (with pre-approved cleanup: touched-file slop, code the change made dead, a typo beside the change), out of scope, integrations (real by default; MOCKED only when asked), assumptions as question and answer, and acceptance criteria with the test that proves each. Pack-owned reference file; copy it with the rest of references/.
• agents/scope-reviewer.md (new) — Opus 5.5 at low effort. Checks a task's diff against the contract: hunks that trace to nothing in it (Warning), criteria with no proving test or a test that passes without the change (Blocker), and mocks or stubs in production paths where the contract says real (Blocker). Copy it in.
• skills/orc/SKILL.md, 3. Make it buildable — new Write the contract step after the ladder: orc writes <scratchpad>/contract.md, states the restatement in one line in chat, and posts the contract on the issue as a Work contract comment. New Contract first (opt-in): with contract first at the start of the invocation, orc asks one approval question before building, takes a change as free text, and builds after at most a second answer. Preflight strips the new opt-in words alongside efficiency mode and, for contract first, stops before touching anything when it can't ask. A batch gets one contract with criteria tagged by issue. What orc does NOT do names the two pre-build questions as the only ones. Refresh.
• skills/orc/SKILL.md, 4 and 5 — tasks are decomposed from the contract's criteria, each criterion owned by one task. Builder dispatches carry the owned criteria with their tests and the contract path. scope-reviewer is a lane on every standard and structural task, and on a trivial task whose files the contract doesn't name. Reviewers get the contract path. Design-brief and design-review dispatches carry the contract and the task's criteria, and the verifier gets the contract when scope-reviewer raised findings. Orc reads each builder's Acceptance line first and sends a test that passed before the change straight back. A scope finding is fixed by reverting the change; one worth keeping becomes its own task, and a contract amendment is recorded in the commit body. Standards lists work-contract.md; Model selection lists scope-reviewer at low effort. Refresh.
• skills/orc/SKILL.md, Cleanup runs and 6. Discoveries — a cleanup run completes its contract after the audit, one In scope line and one criterion per task, with characterization tests exempt from failing-first. Keep the contract current: every task added after the contract (a Small follow-up, a cleanup candidate, a folded-in fix) gets an In scope line tagged follow-up before it is dispatched, so the scope review doesn't flag it. Refresh.
• agents/verifier.md — a confirmed departure from the work contract is a broken documented rule, never a Preference, and not Out of scope merely for sitting in a touched file. Refresh.
• skills/orc/SKILL.md, 9. Report — Landed maps each acceptance criterion to its proving test. Heads-up leads with any MOCKED integration, then the contract's assumptions. Refresh.
• skills/orc/references/builder-contract.md — build mode writes each acceptance test first and confirms it fails for the right reason, then the change; a new rule forbids mocks in production paths unless the contract marks the integration MOCKED; the contract's scope lists are binding; the report gains an Acceptance line (each criterion, its test, before and after). Build steps are renumbered 1 to 10. Refresh; no pack: marker.
• skills/newissue/SKILL.md — new Integrations section (real or mocked), and Done when is now three to six acceptance criteria a test can assert. The self-check covers both. Refresh.
• agents/next-issue-finder.md — new Screen out stale candidates section: two git log calls per candidate (by issue number, and by the paths the issue names) and a code read only on a hit. The output block gains an optional STALE: section, one line per stale issue with the commit and the code evidence. PICK: NONE now also covers a board where every candidate is stale. The agent stays read-only. Refresh.
• skills/orc/SKILL.md, 1. Survey — orc confirms each STALE entry itself, comments the commit and what the code now does, and closes the issue; if the evidence doesn't hold, it leaves the issue alone. Refresh.
• .github/workflows/pack-integrity.yml — the agent-name check also scans skills/shipcheck/SKILL.md, and the reference check covers work-contract.md. Kit CI only.
• INSTALL.md, README.md, CLAUDE.md — list the new skill, agent, and template (23 agents), describe the contract in the run flow, and a 1.11.0 banner. Kit docs only.
• Every skill and agent file — pack: marker moved to orc-pack@1.11.0.

Update steps. Copy the new agents/scope-reviewer.md and skills/orc/references/work-contract.md; orc dispatches the agent and passes the template by name, so an install without them breaks the review loop. Copy skills/shipcheck/SKILL.md to .claude/skills/shipcheck/SKILL.md unless the user declines it (it is optional, like newissue). Refresh agents/next-issue-finder.md, agents/verifier.md, skills/orc/SKILL.md, skills/orc/references/builder-contract.md, and skills/newissue/SKILL.md, and move every pack: marker to orc-pack@1.11.0. Never create or overwrite .claude/required-secrets.md or .claude/smoke-checklist.md; they are repo-owned. No CI, environment, or dependency changes in the target repo. If the install has local edits to next-issue-finder.md, keep the STALE: section of its output block: orc reads it. An install coming from 1.9.x or earlier also applies [1.10.0] and every entry in between.

Breaking changes. No settings or files are removed. Orc runs now post a Work contract comment on the issue they work, standard and structural tasks get one more review lane, and builder reports carry a new Acceptance line. If the install has local edits to builder-contract.md, re-apply them around the renumbered build steps and keep step 4 (tests first) and step 6 (real integrations). Anything that parses the scout's output must tolerate the new optional STALE: section between SET_ASIDE: and GUARDS:.

Files to read. skills/orc/references/work-contract.md, agents/scope-reviewer.md, skills/orc/references/builder-contract.md (Build mode, Scope, Report), skills/shipcheck/SKILL.md, agents/next-issue-finder.md (Screen out stale candidates, Output), and skills/orc/SKILL.md (Preflight, 1. Survey, 3. Make it buildable, 4, 5, Cleanup runs, 6. Discoveries, 9. Report, What orc does NOT do), and agents/verifier.md (Rules).

---

orc-pack v1.9.1

[1.9.1] - 2026-09-28

Summary. A one-paragraph fix to /update-orc. It told the dispatcher to pass model: claude-opus-5-5 to the orc-updater agent, but the Agent tool's model field accepts only a family alias (opus, sonnet, haiku, or fable), so the instruction could not be followed as written. 1.9.0 fixed the same problem in the orc skill and missed this copy. /update-orc now dispatches orc-updater with no model, so the agent's frontmatter pins it to Opus 5.5.

Changed areas.

• skills/update-orc/SKILL.md — step 5 dispatches orc-updater without a model parameter and explains why an opus alias is still avoided. Refresh.
• Every skill and agent file — pack: marker moved to orc-pack@1.9.1.

Update steps. Refresh skills/update-orc/SKILL.md, and move every pack: marker to orc-pack@1.9.1. Nothing else. An install coming from 1.8.0 or earlier also applies the [1.9.0] entry below.

Breaking changes. None.

Files to read. skills/update-orc/SKILL.md (step 5).

---

orc-pack v1.9.0

[1.9.0] - 2026-09-28

Summary. Claude Sonnet 5.5 shipped, and this release fits it into the pack. test-coverage-reviewer moves from Sonnet 5 to Sonnet 5.5: it costs the same, and on CodeRabbit's 13 hard review cases it caught 6 known issues to Sonnet 5's 4 at under half the cost per review. Every other agent keeps its model: at the medium and low effort most of them run, Sonnet 5.5 scores well below Opus 5.5, and the high-effort agents do security work or irreversible data changes, where Opus 5.5 is the safer model. Orc also gains efficiency mode: starting /orc on Sonnet 5.5 asks the user to confirm, then runs with a slightly lower token bill by moving three smaller jobs to Sonnet 5.5, handing more decisions to Opus agents, and skipping the verifier when reviews find only Nits. Starting on Opus 5.5 works as before, with one fix that applies to every run: orc now dispatches through the Agent tool's family aliases, the only values its model field accepts, instead of exact model ids.

Changed areas.

• agents/test-coverage-reviewer.md — model: claude-sonnet-5 becomes model: claude-sonnet-5-5; effort: high is unchanged. Refresh. If the installed copy pins a different model on purpose, keep the local choice and say so in the report.
• skills/orc/SKILL.md, Preflight — sorts the session model into cases: Opus 5.5 and other capable models start straight away, Sonnet 5.5 asks the user to confirm efficiency mode with AskUserQuestion (fixed question text), and Opus 5 is still refused. An invocation that starts with efficiency mode answers the question in advance on Sonnet 5.5 and is ignored, with a note, on other models. Preflight now also checks CLAUDE_CODE_SUBAGENT_MODEL_FORCE before the run, and a preflight stop prints no status line. Refresh.
• skills/orc/SKILL.md, Model selection — corrects how orc picks subagent models. The Agent tool's model takes only opus, sonnet, haiku, or fable, so the old rule to pass exact ids could not be followed. Orc now leaves model off any dispatch whose frontmatter is already right, and passes sonnet only for efficiency-mode moves and fable for a round-three fix. Names Sonnet 5.5 for test-coverage-reviewer. Refresh.
• skills/orc/SKILL.md, Efficiency mode (new subsection) — on a confirmed Sonnet 5.5 run, three dispatches go to Sonnet 5.5: next-issue-finder; the build of a trivial task (not data-implementer; a build that grows past trivial keeps its standard review); and a builder's first fix round when it carries no Blocker and no security-reviewer finding, with quality-reviewer always in the re-review. Every reviewer, the verifier, data-implementer, the first build of standard and structural work, and later fix rounds keep their models. Orc takes the design reviewer's side on a second REVISE, routes ready-check fixes and trivial discovery fixes through builders instead of editing code itself, and skips the verifier when a task's panel returns only Nits with no Cleanup candidate (cleanup audits are always verified). Short pointers in the survey step, design step, builder dispatch, verifier step, fix loop, Discoveries, Ready, and the report's Picked line. Refresh.
• skills/orc/SKILL.md, Untrusted input — one sentence saying timer wakes and subagent completion notices are expected harness events, not untrusted input. Refresh.
• README.md — a 1.9.0 banner (replacing the 1.8.0 one and its broken changelog link), a new Pick a model: default or efficiency mode section, and small mentions of Sonnet 5.5 in the intro and to-do tools note. Kit docs only.
• INSTALL.md — Phase 6 mentions efficiency mode and Sonnet 5.5. Kit docs only.
• CLAUDE.md — the model-assignment section records the Sonnet 5.5 move, the effort-level data behind the tiers (with its source), the Agent tool's alias limit, and the efficiency-mode rules and their main risk. Kit docs only; not installed.
• Every skill and agent file — pack: marker moved to orc-pack@1.9.0.

Update steps. Refresh agents/test-coverage-reviewer.md and skills/orc/SKILL.md, and move every pack: marker to orc-pack@1.9.0. Nothing else: no new files, no CI or environment changes. Sonnet 5.5 must be available to the account running orc; it has its own rate-limit pool, separate from Sonnet 5's.

Breaking changes. None for runs started on Opus 5.5. A run started on Sonnet 5.5 now stops for a one-question confirmation, or stops outright in a non-interactive run, unless the invocation starts with efficiency mode. A scheduled or scripted /orc on Sonnet 5.5 must add those words.

Files to read. skills/orc/SKILL.md (Preflight, Model selection, and Efficiency mode) and agents/test-coverage-reviewer.md.

---

orc-pack v1.8.0

[1.8.0] - 2026-09-28

Summary. Two changes. First, orc gains specialist builders: ui-implementer, api-implementer, and data-implementer join the general implementer, orc assigns each task a builder by its primary surface, and every builder now reads one shared builder-contract.md that holds the rules the old implementer.md carried. A new standards/data.md gives data work a standard to write and review against. Second, orc stops paying to re-cache its whole context after a long subagent wait. The prompt cache lasts about an hour, and orc makes no requests while a subagent runs, so any wait longer than that made the next turn re-cache the full context at about twice the input price (a warm read costs about a tenth). Orc now keeps a background heartbeat timer while agents run, so the main thread touches its cache before it expires.

Changed areas.

• skills/orc/references/builder-contract.md (new) — the rules every builder follows, moved out of agents/implementer.md: what the dispatch contains, untrusted input, reading the standards, brief mode, build mode's core steps, scope, never committing, no suppressions, fix rounds, and the report format (with a new optional Deploy notes heading). Nothing the implementer enforced was dropped or weakened. Every builder reads it first, and each builder's agent file repeats the hard rules (no commits, no suppressions, scope, tests with exact output) in case the contract cannot be read. No pack: marker; it is pack-owned like the rest of references/. Copy in.
• skills/orc/references/standards/data.md (new) — migrations, expand/contract changes, locks on large tables, backfills, indexes, query safety, destructive operations and environments, and the slop signatures of data code. Read by data-implementer (and any builder whose task touches schema, migrations, or queries), design-reviewer on data briefs, and architecture-reviewer when a diff touches schema or migrations. Copy in.
• agents/implementer.md — rewritten as a thin general builder. It reads the contract, then carries judgement for its lane: tooling, config, scripts, docs-adjacent code, cross-cutting changes, and cleanup runs. It stays the fallback for any task orc cannot route. Same model and effort (Opus 5.5, medium). Refresh, then see Update steps for local edits.
• agents/ui-implementer.md (new) — components, pages, layouts, styles, tokens, client state and interaction, and user-facing copy. Always reads standards/ui.md; builds on the repo's design system and tokens, every state, keyboard and screen-reader support, phone width, and every theme; always reports UI surfaces for ui-reviewer. Opus 5.5, medium effort. Copy in.
• agents/api-implementer.md (new) — routes, endpoints, loaders and actions, services, integrations, and background jobs. Boundary validation with the repo's library, consistent error contracts, authorization at the right layer, idempotent retried writes, bounded upstream calls, no secrets or personal data in logs, contract-level tests, and no silent break of external callers. Always reports Deploy notes. Opus 5.5, medium effort. Copy in.
• agents/data-implementer.md (new) — schema changes, migrations, backfills, seed data, ORM models, indexes, and heavy query changes. Uses the repo's migration tool, reversible migrations, expand/contract for live tables, batched and resumable backfills, lock awareness, and indexes matched to real queries. It never runs a migration or backfill against anything but a local or test database, and always reports the deploy order under Deploy notes. Opus 5.5, high effort, because its mistakes are the hardest to reverse. Copy in.
• skills/orc/SKILL.md — step 4 now assigns each task a builder by its primary surface, with a routing table (ui-implementer, api-implementer, data-implementer, and implementer for everything else); a task spanning surfaces is split, or routed to its riskiest surface (data over API over UI). Step 5 dispatches the task's assigned builder in brief mode, build mode, fix rounds (same builder type), and the round-three Fable 5.1 escalation, and still never runs builders in parallel. The Standards table adds builder-contract.md and standards/data.md and names the builders that read each file. Model selection lists every builder.
• skills/orc/SKILL.md — new Waiting on subagents section. Before ending a turn to wait on a subagent, orc starts a background timer of about 50 minutes (for example, sleep 3000). Each wake reads the still-warm cache, re-arms the timer if agents are still running, and ends. Orc stops the timer when the last agent finishes.
• skills/orc/references/standards/code.md, comments.md, structure.md, and ui.md — the "who writes to this" line names the builders instead of "the implementer". Refresh.
• skills/orc/references/design-brief.md — names the task's builder as the author, adds a partial state and a Client state line to the UI section, and adds a Data section (schema change, readers and writers, phases, backfill, locks and indexes, deploy order). Refresh.
• agents/design-reviewer.md — reads data.md for data work and gains a Data checklist for briefs with a Data section. Refresh.
• agents/architecture-reviewer.md — reads data.md when the diff touches schema, migrations, or data queries. Refresh.
• agents/quality-reviewer.md — one-word wording change from "implementer" to "builder". Refresh.
• agents/orc-updater.md — the conflict-handling pointer now names INSTALL.md Phase 3 (it wrongly named UPDATE.md Phase 3). Refresh.
• .github/workflows/pack-integrity.yml (kit CI only, not installed) — check 2 recognizes *-implementer agent names, and check 4 verifies builder-contract.md exists.
• Every skill and agent file — pack: marker moved to orc-pack@1.8.0.

Update steps.

• Copy in the three new agents and the two new reference files. If the repo already has its own agent named ui-implementer, api-implementer, or data-implementer with no pack: marker, route it through INSTALL.md Phase 3 conflict handling.
• Refresh skills/orc/SKILL.md, and re-apply any repo-local edits the install's history shows.
• Local edits to agents/implementer.md need a new home. Most of the old implementer's body now lives in builder-contract.md. For each local edit the install's history shows on implementer.md, decide where it belongs: an edit to a rule every builder follows (the dispatch, standards, brief or build steps, scope, commits, suppressions, fix rounds, the report) goes into .claude/skills/orc/references/builder-contract.md, so every builder gets it; an edit specific to one kind of work goes into that builder's agent file; an edit about tooling, config, scripts, or cleanup stays in implementer.md. Report each edit and where you put it.
• Update the pack: marker on every agent and skill file to orc-pack@1.8.0.
• The heartbeat needs the main session to run a background shell command that wakes it on exit. Claude Code's Bash tool with run_in_background does this. No re-vet, CI, or environment change is required.

Breaking changes. None for a clean install. An install with local edits to agents/implementer.md must re-home them as described in Update steps, or those edits stop applying to UI, API, and data tasks. Long runs show a short heartbeat turn about every 50 minutes while agents work.

Files to read. skills/orc/references/builder-contract.md, agents/implementer.md, agents/ui-implementer.md, agents/api-implementer.md, agents/data-implementer.md, skills/orc/references/standards/data.md, and skills/orc/SKILL.md (step 4's builder routing, step 5, Standards, Model selection, and Waiting on subagents).

---

orc-pack v1.7.0

[1.7.0] - 2026-09-22

Summary. Raises orc's quality bar and makes it cheaper to hold. A new set of shared standards files (code, comments, structure, and UI) plus detection-gated framework and platform playbooks (Next.js, Nuxt, SvelteKit, Astro, Cloudflare, Vercel) now live beside the orc skill. The implementer writes to them and every reviewer checks against them. Three new reviewers join the panel: design-reviewer approves a design brief before code is written for structural tasks, comment-reviewer applies a cold-read test to every comment in touched files, and ui-reviewer checks UI changes against the design system and, when the app runs locally, reviews screenshots of the rendered screens. Orc now sizes each task (trivial, standard, or structural) and dispatches only the reviewers the task's size and surfaces call for. The verifier gains a "preference" verdict so reviewers' taste never costs a fix round. A new cleanup mode (/orc clean up the slop in <path>) runs behavior-preserving refactors of slop code and comments. quality-reviewer and architecture-reviewer move to Opus 5.5.

Changed areas.

• skills/orc/references/ (new directory) — the standards orc passes to its agents: standards/code.md, standards/comments.md, standards/structure.md, standards/ui.md, frameworks/next.md, frameworks/nuxt.md, frameworks/sveltekit.md, frameworks/astro.md, platforms/cloudflare.md, platforms/vercel.md, and the design-brief.md template. Copy the whole directory to <root>/.claude/skills/orc/references/. These files carry no pack: marker; they belong to the orc skill directory, which is pack-owned.
• agents/design-reviewer.md (new) — reviews a structural task's design brief before implementation. Opus 5.5, medium effort. Copy in.
• agents/comment-reviewer.md (new) — the cold-read test, JSDoc on exports, and slop comments, across every comment in a touched file. Opus 5.5, low effort. Copy in.
• agents/ui-reviewer.md (new) — design system, hierarchy, states, interaction, responsive, accessibility, and copy, plus an optional visual pass through the repo's own Playwright. It never installs packages or browsers. Opus 5.5, medium effort. Copy in.
• agents/quality-reviewer.md — rewritten around readability and the AI slop code signatures. Accessibility and design-system checks move to ui-reviewer, comments move to comment-reviewer, and the stale reference to a non-existent "duplication reviewer" is gone. model: claude-opus-5-5, effort: medium (was claude-sonnet-5, high). Refresh.
• agents/architecture-reviewer.md — adds separation of concerns, file and folder structure, and design-brief conformance. model: claude-opus-5-5, effort: medium (was claude-sonnet-5, high). Refresh.
• agents/implementer.md — gains a brief mode that writes a design brief without touching the repo, reads the standards before building, cleans up slop in the files it touches (behavior-preserving), and reports UI surfaces and cleanup candidates. Refresh.
• agents/verifier.md — new 🎨 Preference and 🧭 Out-of-scope verdicts, and a "does the fix make the code genuinely better" check. Refresh.
• agents/security-reviewer.md, agents/test-coverage-reviewer.md, agents/docs-writer.md — read the playbooks or standards they are given. The test-coverage reviewer gains a cleanup mode that pins current behavior with characterization tests. Refresh.
• Review scope for quality-reviewer, architecture-reviewer, and comment-reviewer — pre-existing slop in a file the diff touches is now in scope and fixed in the task, tagged (touched file). A touched file that needs more than the task can fix proportionately is raised once as a Cleanup candidate and becomes its own task. Quality debt in untouched files is a cleanup target, listed in the report for a later run. comment-reviewer requires JSDoc only on exports the diff adds or changes; older undocumented exports are cleanup work.
• skills/orc/SKILL.md:
  • New Standards section: how orc finds references/ (the skill's base directory, then a glob fallback) and which files go to which agent.
  • Step 1 detects the framework, deploy target, and UI to resolve the standards set.
  • Step 4 sizes every task: trivial, standard, or structural.
  • Step 5 adds the design-brief step for structural tasks, replaces reviewer selection with a size and surface routing table, skips the verifier when the panel returns no findings, runs fix rounds on Blockers or Warnings (never on Nits alone), and requires commit subjects that name the change rather than the review process.
  • New Cleanup runs section and a Directed — cleanup mode.
  • Discoveries no longer chases quality debt in untouched files; those files are listed under Your call as cleanup targets.
  • Model selection updated to the new tiers, and a new No make-work rule.
  • Commits stage exactly the task's files with git add -- <files> instead of relying on whatever is staged.
  • Fixes the review diff recipe. It was git diff <base> HEAD, which is empty because the implementer never commits; it is now git add --intent-to-add . followed by git diff <base>, so new files appear too.
• .github/workflows/pack-integrity.yml — new check 4: every standards file and playbook the orc skill names must exist under skills/orc/references/.
• INSTALL.md, UPDATE.md, skills/update-orc/SKILL.md, agents/orc-updater.md — the new reference files carry no pack: marker, so every install and update procedure now treats files under skills/orc/references/ as pack-owned: refreshed from the kit, local edits re-applied, and dropped files removed. Without this, a later update would route them into conflict handling as repo-owned files.
• README.md, CLAUDE.md — layout, agent table, standards, and model tiers updated.

Update steps.

• Copy the new skills/orc/references/ directory in full, and the three new agent files.
• Refresh every other agent file and skills/orc/SKILL.md. If a repo-local edit changed quality-reviewer or architecture-reviewer's model: or effort: lines, keep the local choice.
• A repo that wants its own house rules to beat the pack's standards needs no change: every standards file already defers to the repo's CLAUDE.md/AGENTS.md and established patterns. To tune the standards for one repo, edit that install's copies under .claude/skills/orc/references/ and record the edit in provenance.
• ui-reviewer's visual pass uses Playwright only when the repo already has it with browsers installed. Nothing new is installed or vetted.
• No re-vet, CI, or environment change is required in the target repo.

Breaking changes. None to invocation or outcomes. Behavior changes to expect:

• Reviews are stricter on readability, comments, and structure, and pre-existing slop in touched files is now fixed in the task, so diffs can be larger than before.
• Fix rounds now run on verified Warnings as well as Blockers.
• A cleanup audit that finds nothing ends FINISHED (no build).
• Structural tasks add two dispatches (a brief and its review) before implementation.
• Cost moves up for the two reviewers that moved to Opus 5.5, and down for trivial tasks, which now get a single reviewer and skip the verifier when it finds nothing.

Files to read. skills/orc/SKILL.md (Standards, step 1, steps 4 and 5, Cleanup runs, Discoveries, and Model selection), every file under skills/orc/references/, UPDATE.md, agents/implementer.md, agents/design-reviewer.md, agents/comment-reviewer.md, agents/ui-reviewer.md, agents/quality-reviewer.md, agents/architecture-reviewer.md, and agents/verifier.md.

---

orc-pack v1.6.1

[1.6.1] - 2026-09-22

Summary. Moves the three routine reviewers back to Sonnet 5 and lowers the verifier's effort. Security review and verification stay on Opus 5.5. The routine reviewers only need to not miss real problems, because the Opus 5.5 verifier filters out their false positives. Keeping them at high effort on Sonnet 5 costs less per task without weakening the security path.

Changed areas.

• agents/architecture-reviewer.md, agents/quality-reviewer.md, agents/test-coverage-reviewer.md — model: claude-sonnet-5, effort: high (were claude-opus-5-5, medium). Refresh.
• agents/verifier.md — effort: medium (was high); model unchanged (claude-opus-5-5). Refresh.
• skills/orc/SKILL.md — Model selection now lists the routine reviewers on Sonnet 5 and requires that security review and verification always stay on Opus 5.5. Refresh.
• CLAUDE.md, README.md — tier descriptions updated to match. Refresh README.md if the install carries it.

Update steps.

• Refresh the four agent files and skills/orc/SKILL.md. If a repo-local edit changed one of those agents' model: or effort: lines, keep the local choice.
• Sonnet 5 (claude-sonnet-5) must be available to the account or platform.
• No re-vet, CI, or environment change is required.

Breaking changes. None.

Files to read. skills/orc/SKILL.md (Model selection), CLAUDE.md (Model assignments are intentional), and the frontmatter of the four changed agents.

---

orc-pack v1.5.0

[1.5.0] - 2026-09-21

Summary. Removes the CodeGraph integration from the pack. CodeGraph proved unreliable in practice — it caused failures and its structural answers were not dependable enough to justify the agent, the CLI install, and the supply-chain vet it carried. orc no longer dispatches a codegraph agent, the pack no longer installs or vets the CLI, and a fresh install has nothing to do with CodeGraph. fallow is unchanged and stays a default-installed, vetted tool.

Changed areas.

• agents/codegraph.md (removed) — the agent is gone. Delete <root>/.claude/agents/codegraph.md from an existing install.
• skills/orc/SKILL.md — codegraph removed from the review panel, from step 7's affected-tests scoping, and from the Haiku tooling-runner list. Refresh.
• agents/dependency-vetter.md — no longer names CodeGraph or its known-good record; the vet now covers fallow (and any future third-party tool) on the same install-latest-after-vet policy. Refresh.
• agents/orc-updater.md — the update-steps bullet now says "re-vet fallow if named" rather than CodeGraph/fallow. Refresh.
• .claude-plugin/codegraph.known-good.json (removed) — the install-latest-after-vet record for CodeGraph is gone. It is kit config that was never copied into an install, so an install has nothing to delete.
• .github/workflows/codegraph-vet.yml (removed) — the release canary for CodeGraph is gone. Only relevant if the repo carries the pack's own CI.
• .github/workflows/pack-integrity.yml — codegraph dropped from the known agent-name list. Refresh if the repo carries this CI.
• README.md, INSTALL.md, UPDATE.md, CLAUDE.md — CodeGraph documentation, install phase, and update phase removed; fallow is the remaining third-party tool. Refresh.

Update steps.

• Delete the CodeGraph agent from the install: <root>/.claude/agents/codegraph.md. Do not leave it behind — orc no longer dispatches it, so a stale copy is dead weight that misleads the next reader.
• Remove any CodeGraph entry from <root>/.claude/orc-pack.provenance.md (installed version, integrity hash, vet verdict).
• Nothing installs CodeGraph going forward: do not run its install or its vet again. If the CLI is already on the machine, the pack no longer uses it — leaving or uninstalling it is the user's call.
• Re-vet fallow if the install has it, per the revised UPDATE.md Phase 3.

Breaking changes. None at runtime. An install that deletes agents/codegraph.md and refreshes the skills is consistent with the target kit.

Files to read before applying. UPDATE.md (Phase 2 and the revised Phase 3), INSTALL.md (Phase 2 mapping and the revised Phase 4.5), skills/orc/SKILL.md, this entry.

---