Ship the client MVP as 0.10.0 and the rest of the pluggability work as 0.11.0, and move the leftover 0.9.0 items to 0.12.0
CI / tests (pull_request) Successful in 2m13s

This commit was merged in pull request #515.
This commit is contained in:
2026-11-12 22:43:26 +01:00
parent 0c98931df0
commit 525f9ccb9d
+126 -126
View File
@@ -12,39 +12,39 @@
| ID | Release | Blocked by | Exempt | Item | L | E | I | R | S | Moves | Score |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 319 | 0.10.0 | | | **Start a fresh session with a brief of the chunk when a rung step changes the worker's model, in place of resuming the old session.** | 3 | 2 | 3 | 6 | 4 | 1.1 by 2, 1.4 by 2 | 1.2 |
| 309 | 0.10.0 | | | **Run a scoring run's second pass on one head with two seats in place of four, once `larv-review agreement` shows the seats agree with each other more than a run agrees with its rerun.** | 2 | 2 | 2 | 4 | 2 | 1.4 by 3 | 1.2 |
| 299 | 0.10.0 | | | **Replay the archived judge verdicts through code rules and a classifier trained locally on the archive, and file an item for each verdict kind one of them decides with no missed interrupt.** | 2 | 2 | 2 | 4 | 3 | 1.4 by 3 | 1.0 |
| 311 | 0.10.0 | | | **Hand each worker's claude a Bash output ceiling that saves longer output to a file, and refuse a Read of a session transcript under `~/.claude/projects`.** | 2 | 2 | 3 | 5 | 3 | 1.4 by 3 | 1.0 |
| 316 | 0.10.0 | | | **Launch comprehension seats as an agent type of their own under `agents/`, without `CLAUDE.md` and the skill and agent listings a general-purpose subagent carries.** | 2 | 2 | 2 | 4 | 3 | 1.4 by 3 | 1.0 |
| 339 | 0.10.0 | | | **Lower the reading cost of the files a chunk reads most before its first commit.** | 2 | 2 | 1 | 3 | 3 | 1.4 by 3 | 1.0 |
| 341 | 0.10.0 | | | **Hand each worker's claude a lower Read limit through `CLAUDE_CODE_FILE_READ_MAX_OUTPUT_TOKENS`, kept only where replays of the archived review rounds find what two unchanged replays find.** | 2 | 2 | 3 | 5 | 3 | 1.4 by 3 | 1.0 |
| 310 | 0.10.0 | | | **Compact a waiting worker's session shortly before its prompt cache expires, only when the human has not answered an interrupt by then.** | 3 | 2 | 3 | 6 | 4 | 1.4 by 3 | 0.9 |
| 298 | 0.10.0 | | | **Send the judge only the decision entries and the transcript tail a verdict needs, and keep the cut only where `judge-replay` changes no verdict kind.** | 3 | 3 | 3 | 7 | 4 | 1.4 by 2, 1.5 by 1 | 0.9 |
| 331 | 0.10.0 | | | **Reword four decisions to expire when the rungs their replays tested change, not whenever a model joins the ladder.** | 1 | 1 | 1 | 1 | 2 | 1.4 by 2 | 0.8 |
| 327 | 0.10.0 | | | **Have `larv-review` print a round commit's title prefix, counted from the round's report, and have rule-hook refuse a round title whose counts differ from it.** | 3 | 2 | 3 | 6 | 4 | 1.1 by 1, 1.4 by 1 | 0.6 |
| 343 | 0.11.0 | | | **Read every setting a project needs from one explicit config in the project's own git repo, `~/.claude-projects/<key>/`, found from the checkout's path, and fail naming each setting the config leaves out.** | 3 | 2 | 3 | 6 | 4 | 1.6 by 6 | 1.6 |
| 347 | 0.11.0 | 343 | | **Hand reviewers the technical principles from the folder a project's config names, in place of `~/.claude/principles/`.** | 2 | 2 | 3 | 5 | 2 | 1.6 by 6 | 2.2 |
| 346 | 0.11.0 | 343 | | **Read a project's goals and audience from the file its config names, in place of `README.md`'s `## Goals` and `## Audience`.** | 2 | 2 | 3 | 5 | 3 | 1.6 by 6 | 2.0 |
| 344 | 0.11.0 | 343 | | **Read and write the plan, `AGENTS.md`, the records and the changelog at the paths a project's config names, in every tool, prompt, rule and skill.** | 4 | 2 | 3 | 7 | 6 | 1.6 by 7 | 0.9 |
| 345 | 0.11.0 | 344 | | **Commit each change a chunk makes to a plan or record outside the checkout into the project's `~/.claude-projects/<key>/` repo, titled like the chunk's own commits.** | 3 | 2 | 3 | 6 | 4 | 1.6 by 4 | 1.1 |
| 348 | 0.12.0 | 343 | | **Read each project's credentials from the env file its config names, fail naming each missing key, and keep every value out of prompts, transcripts, logs and the archive.** | 2 | 2 | 4 | 6 | 3 | 1.6 by 2, 4 by 5 | 6.9 |
| 354 | 0.12.0 | 344 | | **Read and write the decisions, score anchors, comprehension history, reading cost and changelog each through the backend its project's config names, with markdown files as one backend.** | 3 | 2 | 3 | 6 | 6 | 1.6 by 4, code health by 1 | 1.2 |
| 364 | 0.12.0 | 354 | | **Read a client repo's own `CLAUDE.md` and `AGENTS.md`, propose to its team each change that aligns them with this setup, and record each conflict left as a decision in the project's records.** | 2 | 2 | 2 | 4 | 2 | 1.6 by 4, 1.2 by 2 | 2.2 |
| 360 | 0.12.0 | 343 | | **Make `chunk-pr merge` wait for the human approvals a project's config requires, as part of the gate, before it merges.** | 2 | 2 | 3 | 5 | 3 | 1.6 by 5 | 1.6 |
| 357 | 0.12.0 | 343 | | **Write prose in the language a project's config names, and title its commits, pull requests and branches by the templates it names.** | 3 | 2 | 3 | 6 | 4 | 1.6 by 5 | 1.4 |
| 350 | 0.12.0 | 344 | | **Route every read and write of the plan, by tools and by agents, through one plan-store interface and `todo-item`, with the markdown file as one backend.** | 4 | 2 | 3 | 7 | 7 | 1.6 by 4, code health by 2 | 1.1 |
| 349 | 0.12.0 | | | **Convert markdown to ADF and back with `@larvit/adf-codec`'s portable calls, through a pinned Node shim run in a container with a memory limit.** | 2 | 1 | 2 | 3 | 2 | 1.6 by 3 | 1.1 |
| 351 | 0.12.0 | 348, 349, 350 | | **Read a project's plan from the child issues of a Jira Cloud epic its config names, ranked in Jira's backlog order.** | 3 | 2 | 3 | 6 | 5 | 1.6 by 6 | 1.2 |
| 353 | 0.12.0 | 346, 351 | | **Let a project's config take its goals and audience from a Jira epic's description, under the same two headings.** | 2 | 2 | 3 | 5 | 3 | 1.6 by 4 | 1.3 |
| 358 | 0.12.0 | 343 | | **Let a project's config allow its runtime and tests on the host, and set the comprehension floor that marks code health very important.** | 2 | 2 | 3 | 5 | 3 | 1.6 by 4 | 1.3 |
| 362 | 0.12.0 | 348 | | **Send loop-claude's notices through the notifier a project's config names, with Mattermost as the first backend.** | 2 | 1 | 2 | 3 | 3 | 1.6 by 3 | 1.0 |
| 363 | 0.12.0 | 362 | | **Add a Slack backend to the notifier, sending each notice as a DM from a bot.** | 2 | 1 | 2 | 3 | 3 | 1.6 by 4 | 1.3 |
| 352 | 0.12.0 | 351 | | **Write a project's plan to its Jira epic: file items as child issues, edit those the loop filed, score unscored issues with a comment, and move a finished item to the status its config names.** | 3 | 2 | 3 | 6 | 5 | 1.6 by 6 | 1.2 |
| 355 | 0.12.0 | 348, 349, 354 | | **Store a record on the Confluence Cloud page its config names, as markdown converted through the adf-codec shim, and interrupt the human when a teammate saved the page since the loop read it.** | 3 | 2 | 3 | 6 | 5 | 1.6 by 6 | 1.2 |
| 356 | 0.12.0 | 347 | | **Hand reviewers the review standards from the folder a project's config names, in place of `~/.claude/skills/larv-review/standards/`.** | 2 | 2 | 3 | 5 | 2 | 1.6 by 3 | 1.1 |
| 359 | 0.12.0 | 350 | | **Tie a project's releases to what its config names: semver tags, Jira Fix Versions, sprints, or none.** | 3 | 2 | 3 | 6 | 4 | 1.6 by 3 | 0.8 |
| 361 | 0.12.0 | 360 | | **Start the next item that doesn't touch a waiting pull request's files, in its own worktree and session, while that pull request waits for a human approval, and answer review comments on waiting pull requests as they arrive.** | 4 | 2 | 3 | 7 | 7 | 1.6 by 4, 1.5 by 5 | 0.7 |
| 343 | 0.10.0 | | | **Read every setting a project needs from one explicit config in the project's own git repo, `~/.claude-projects/<key>/`, found from the checkout's path, and fail naming each setting the config leaves out.** | 3 | 2 | 3 | 6 | 4 | 1.6 by 6 | 1.6 |
| 347 | 0.10.0 | 343 | | **Hand reviewers the technical principles from the folder a project's config names, in place of `~/.claude/principles/`.** | 2 | 2 | 3 | 5 | 2 | 1.6 by 6 | 2.2 |
| 346 | 0.10.0 | 343 | | **Read a project's goals and audience from the file its config names, in place of `README.md`'s `## Goals` and `## Audience`.** | 2 | 2 | 3 | 5 | 3 | 1.6 by 6 | 2.0 |
| 344 | 0.10.0 | 343 | | **Read and write the plan, `AGENTS.md`, the records and the changelog at the paths a project's config names, in every tool, prompt, rule and skill.** | 4 | 2 | 3 | 7 | 6 | 1.6 by 7 | 0.9 |
| 345 | 0.10.0 | 344 | | **Commit each change a chunk makes to a plan or record outside the checkout into the project's `~/.claude-projects/<key>/` repo, titled like the chunk's own commits.** | 3 | 2 | 3 | 6 | 4 | 1.6 by 4 | 1.1 |
| 348 | 0.11.0 | 343 | | **Read each project's credentials from the env file its config names, fail naming each missing key, and keep every value out of prompts, transcripts, logs and the archive.** | 2 | 2 | 4 | 6 | 3 | 1.6 by 2, 4 by 5 | 6.9 |
| 354 | 0.11.0 | 344 | | **Read and write the decisions, score anchors, comprehension history, reading cost and changelog each through the backend its project's config names, with markdown files as one backend.** | 3 | 2 | 3 | 6 | 6 | 1.6 by 4, code health by 1 | 1.2 |
| 364 | 0.11.0 | 354 | | **Read a client repo's own `CLAUDE.md` and `AGENTS.md`, propose to its team each change that aligns them with this setup, and record each conflict left as a decision in the project's records.** | 2 | 2 | 2 | 4 | 2 | 1.6 by 4, 1.2 by 2 | 2.2 |
| 360 | 0.11.0 | 343 | | **Make `chunk-pr merge` wait for the human approvals a project's config requires, as part of the gate, before it merges.** | 2 | 2 | 3 | 5 | 3 | 1.6 by 5 | 1.6 |
| 357 | 0.11.0 | 343 | | **Write prose in the language a project's config names, and title its commits, pull requests and branches by the templates it names.** | 3 | 2 | 3 | 6 | 4 | 1.6 by 5 | 1.4 |
| 350 | 0.11.0 | 344 | | **Route every read and write of the plan, by tools and by agents, through one plan-store interface and `todo-item`, with the markdown file as one backend.** | 4 | 2 | 3 | 7 | 7 | 1.6 by 4, code health by 2 | 1.1 |
| 349 | 0.11.0 | | | **Convert markdown to ADF and back with `@larvit/adf-codec`'s portable calls, through a pinned Node shim run in a container with a memory limit.** | 2 | 1 | 2 | 3 | 2 | 1.6 by 3 | 1.1 |
| 351 | 0.11.0 | 348, 349, 350 | | **Read a project's plan from the child issues of a Jira Cloud epic its config names, ranked in Jira's backlog order.** | 3 | 2 | 3 | 6 | 5 | 1.6 by 6 | 1.2 |
| 353 | 0.11.0 | 346, 351 | | **Let a project's config take its goals and audience from a Jira epic's description, under the same two headings.** | 2 | 2 | 3 | 5 | 3 | 1.6 by 4 | 1.3 |
| 358 | 0.11.0 | 343 | | **Let a project's config allow its runtime and tests on the host, and set the comprehension floor that marks code health very important.** | 2 | 2 | 3 | 5 | 3 | 1.6 by 4 | 1.3 |
| 362 | 0.11.0 | 348 | | **Send loop-claude's notices through the notifier a project's config names, with Mattermost as the first backend.** | 2 | 1 | 2 | 3 | 3 | 1.6 by 3 | 1.0 |
| 363 | 0.11.0 | 362 | | **Add a Slack backend to the notifier, sending each notice as a DM from a bot.** | 2 | 1 | 2 | 3 | 3 | 1.6 by 4 | 1.3 |
| 352 | 0.11.0 | 351 | | **Write a project's plan to its Jira epic: file items as child issues, edit those the loop filed, score unscored issues with a comment, and move a finished item to the status its config names.** | 3 | 2 | 3 | 6 | 5 | 1.6 by 6 | 1.2 |
| 355 | 0.11.0 | 348, 349, 354 | | **Store a record on the Confluence Cloud page its config names, as markdown converted through the adf-codec shim, and interrupt the human when a teammate saved the page since the loop read it.** | 3 | 2 | 3 | 6 | 5 | 1.6 by 6 | 1.2 |
| 356 | 0.11.0 | 347 | | **Hand reviewers the review standards from the folder a project's config names, in place of `~/.claude/skills/larv-review/standards/`.** | 2 | 2 | 3 | 5 | 2 | 1.6 by 3 | 1.1 |
| 359 | 0.11.0 | 350 | | **Tie a project's releases to what its config names: semver tags, Jira Fix Versions, sprints, or none.** | 3 | 2 | 3 | 6 | 4 | 1.6 by 3 | 0.8 |
| 361 | 0.11.0 | 360 | | **Start the next item that doesn't touch a waiting pull request's files, in its own worktree and session, while that pull request waits for a human approval, and answer review comments on waiting pull requests as they arrive.** | 4 | 2 | 3 | 7 | 7 | 1.6 by 4, 1.5 by 5 | 0.7 |
| 319 | 0.12.0 | | | **Start a fresh session with a brief of the chunk when a rung step changes the worker's model, in place of resuming the old session.** | 3 | 2 | 3 | 6 | 4 | 1.1 by 2, 1.4 by 2 | 1.2 |
| 309 | 0.12.0 | | | **Run a scoring run's second pass on one head with two seats in place of four, once `larv-review agreement` shows the seats agree with each other more than a run agrees with its rerun.** | 2 | 2 | 2 | 4 | 2 | 1.4 by 3 | 1.2 |
| 299 | 0.12.0 | | | **Replay the archived judge verdicts through code rules and a classifier trained locally on the archive, and file an item for each verdict kind one of them decides with no missed interrupt.** | 2 | 2 | 2 | 4 | 3 | 1.4 by 3 | 1.0 |
| 311 | 0.12.0 | | | **Hand each worker's claude a Bash output ceiling that saves longer output to a file, and refuse a Read of a session transcript under `~/.claude/projects`.** | 2 | 2 | 3 | 5 | 3 | 1.4 by 3 | 1.0 |
| 316 | 0.12.0 | | | **Launch comprehension seats as an agent type of their own under `agents/`, without `CLAUDE.md` and the skill and agent listings a general-purpose subagent carries.** | 2 | 2 | 2 | 4 | 3 | 1.4 by 3 | 1.0 |
| 339 | 0.12.0 | | | **Lower the reading cost of the files a chunk reads most before its first commit.** | 2 | 2 | 1 | 3 | 3 | 1.4 by 3 | 1.0 |
| 341 | 0.12.0 | | | **Hand each worker's claude a lower Read limit through `CLAUDE_CODE_FILE_READ_MAX_OUTPUT_TOKENS`, kept only where replays of the archived review rounds find what two unchanged replays find.** | 2 | 2 | 3 | 5 | 3 | 1.4 by 3 | 1.0 |
| 310 | 0.12.0 | | | **Compact a waiting worker's session shortly before its prompt cache expires, only when the human has not answered an interrupt by then.** | 3 | 2 | 3 | 6 | 4 | 1.4 by 3 | 0.9 |
| 298 | 0.12.0 | | | **Send the judge only the decision entries and the transcript tail a verdict needs, and keep the cut only where `judge-replay` changes no verdict kind.** | 3 | 3 | 3 | 7 | 4 | 1.4 by 2, 1.5 by 1 | 0.9 |
| 331 | 0.12.0 | | | **Reword four decisions to expire when the rungs their replays tested change, not whenever a model joins the ladder.** | 1 | 1 | 1 | 1 | 2 | 1.4 by 2 | 0.8 |
| 327 | 0.12.0 | | | **Have `larv-review` print a round commit's title prefix, counted from the round's report, and have rule-hook refuse a round title whose counts differ from it.** | 3 | 2 | 3 | 6 | 4 | 1.1 by 1, 1.4 by 1 | 0.6 |
| 365 | 0.13.0 | 345, 352 | | **Have `score-outcomes` read the finished and dropped items of a plan outside the checkout, from the project repo's history or from Jira's issue changelogs.** | 3 | 2 | 2 | 5 | 4 | 1.6 by 2, 1.1 by 2 | 1.2 |
| 366 | 0.13.0 | 345, 352 | | **Count the chunks of a project whose plan sits outside the checkout in the routing figures, from the project repo's history or from Jira.** | 3 | 2 | 2 | 5 | 4 | 1.6 by 2, 1.4 by 2 | 1.1 |
| 367 | 0.13.0 | 352 | | **Check that every write to a Jira plan leaves it ranked, as rule-hook's `todo-rank-check` checks `todo.md` today.** | 2 | 2 | 3 | 5 | 3 | 1.6 by 2 | 0.7 |
@@ -91,6 +91,98 @@
## Details
### 343. Read every setting a project needs from one explicit config in the project's own git repo, `~/.claude-projects/<key>/`, found from the checkout's path, and fail naming each setting the config leaves out.
lilleman's calls, 2026-11-12. 0.10.0 is the bare minimum to run loop-claude on a client's legacy project: the plan, the records, the goals and audience and the principles move to paths the config names, and nothing else becomes pluggable yet. Each project has a folder `~/.claude-projects/<key>/`, where `<key>` is the project's name, and the folder is a git repo of its own. It holds the config, and for a client project the files the config points at. The config names the checkout it belongs to, so a tool run inside a checkout finds its config. A project is always a whole git checkout, never a subfolder.
No setting has a built-in default: a missing config, or a missing setting in one, fails and names what is missing. 0.10.0's settings are the paths of `todo.md`, `AGENTS.md`, `docs/decisions.md`, `docs/score-anchors.md`, `docs/comprehension-history.md`, `docs/reading-cost.md` and `CHANGELOG.md`, the goals-and-audience file (item 346), and the principles folder (item 347). A path may point into the checkout or into the project's folder. This chunk also writes a config for every project under `~/code` that names the files each one uses today, so each keeps running unchanged. The config's spelling is a syntax lilleman writes, so a panel seated as lilleman picks it before it ships. README → Install says how to set up a client project, on one sandbox VM per client.
### 347. Hand reviewers the technical principles from the folder a project's config names, in place of `~/.claude/principles/`.
lilleman's call, 2026-11-12: every project must set this folder. It replaces `~/.claude/principles/` and doesn't add to it, so the projects under `~/code` name `~/.claude/principles/`. The `Applies to` line works inside the named folder. Readers: larv-review's principles pass, the reviewers under `agents/`, and `CLAUDE.md` and the skills wherever they name `~/.claude/principles/`.
### 346. Read a project's goals and audience from the file its config names, in place of `README.md`'s `## Goals` and `## Audience`.
lilleman's call, 2026-11-12: for a client project, goals and audience live in a file of their own in `~/.claude-projects/<key>/`, with the same `## Goals` and `## Audience` headings and the same rules (`rules/goals-and-audience.md`), which must then load for that file too. A project reads only its own premises, never a parent project's. Every reader uses the configured file: loop-claude's prompts and judge, `rank_todo.py`'s goal weights and Moves check, larv-review's principles pass, the comprehension panel, and the places in `CLAUDE.md` that say `README.md`. A legacy project whose comprehension score would pull its plan towards refactoring can void code health in that file with the `**Void:**` line, as a README can today.
### 344. Read and write the plan, `AGENTS.md`, the records and the changelog at the paths a project's config names, in every tool, prompt, rule and skill.
Today about 40 modules, and `CLAUDE.md`, `rules/` and the skills, name `todo.md`, `AGENTS.md`, `docs/decisions.md`, `docs/score-anchors.md`, `docs/comprehension-history.md`, `docs/reading-cost.md` and `CHANGELOG.md` as files at the checkout's root. Among them are `rank_todo.py`, `todo-item`, `chunk-preflight`, loop-claude's routing, chunk items, judge records and prompts, larv-review, loop-inbox, `release_plan`, `subagent-rung` and `reviewer_context.principles_line`. After this item each of them uses the configured path. Agents get the path in the prompts and in the output of `todo-item` and `chunk-preflight`. Keep each file's name, so `rules/todo-items.md` and `rules/decisions.md` still load by their `paths:` patterns wherever the file lives.
lilleman's call, 2026-11-12: `AGENTS.md`'s path is configurable too. The tooling treats `AGENTS.md` as ours, since it lists the project principles and the decision titles, and a client repo may hold its own `AGENTS.md`, which stays theirs. Where the plan sits outside the checkout, loop-claude reads the chunk's items from that file, not from `todo.md` on the default branch. Tools that read `todo.md`'s git history in the checkout, `score-outcomes` and the routing figures, say they don't support a plan outside the checkout, and the loop goes on, until 0.13.0. R is above 5, so this item ships alone.
### 345. Commit each change a chunk makes to a plan or record outside the checkout into the project's `~/.claude-projects/<key>/` repo, titled like the chunk's own commits.
lilleman's call, 2026-11-12: each `~/.claude-projects/<key>/` is a git repo of its own, so plan and record edits keep a history. Those edits don't ride the chunk's branch: one loop runs per project, and it commits to the project repo's only branch. rule-hook's rank check, and the `Drop item` title rule, apply to these commits as they do in the checkout. Whether these repos are pushed anywhere for backup is a call for later: client data stays on the client's VM (goal 4).
### 348. Read each project's credentials from the env file its config names, fail naming each missing key, and keep every value out of prompts, transcripts, logs and the archive.
lilleman's call, 2026-11-12: credentials (Jira and Confluence API token, GitHub, Mattermost or Slack token) come from a per-project env file such as `~/.claude-projects/<key>.env`, outside the project's repo and named in the config. Which account each service uses is a config setting: a bot account or lilleman's own. A test shows a value never reaches a worker or judge prompt, a log line or the transcript archive, which guards goal 4. Today `~/.claude/.env` serves every project.
### 354. Read and write the decisions, score anchors, comprehension history, reading cost and changelog each through the backend its project's config names, with markdown files as one backend.
lilleman's call, 2026-11-12: each record has its own backend setting, so a client project can keep, say, decisions and changelog in Confluence and anchors and comprehension history as files. Which goes where is asked per client at setup. Every tool and agent that reads or writes a record goes through the interface: the judge's records, `rank_todo.py`'s decision check, `todo-item finish --changelog`, larv-review's scoring and places, `score-outcomes`, `rules/decisions.md`. The plan store's contract suite has a twin for records, on fakes. R is above 5, so this item ships alone.
### 364. Read a client repo's own `CLAUDE.md` and `AGENTS.md`, propose to its team each change that aligns them with this setup, and record each conflict left as a decision in the project's records.
lilleman's call, 2026-11-12: the loop may change every file in a client repo, but in agreement with the team, so a change to their instruction files goes to them as a proposal in a pull request. Where a conflict stays, the loop records which rule it follows and why as a decision in the project's decision record, so it is never asked twice. This is a working rule in `CLAUDE.md`.
### 360. Make `chunk-pr merge` wait for the human approvals a project's config requires, as part of the gate, before it merges.
lilleman's call, 2026-11-12: this client requires another human's approval on every pull request, so the gate covers it, and the loop still merges once the gate passes. Until this ships, lilleman merges such pull requests by hand and restarts the loop. On GitHub, read the approvals the branch protection asks for; this client's forge is GitHub.
### 357. Write prose in the language a project's config names, and title its commits, pull requests and branches by the templates it names.
lilleman's call, 2026-11-12: a client may want docs and Jira text in another language than English, and may use a commit, PR or branch format other than `CLAUDE.md`'s `ABC-123 - ...` and `ABC-123`. Everything reads the project's setting: `CLAUDE.md` → "Prose" and "Git, commits, and PRs", `chunk-preflight`'s branch and worktree names, `chunk-pr`'s titles, rule-hook's round-title checks and `score-outcomes`' `Drop item` titles. The prose reviewer judges prose in the configured language.
### 350. Route every read and write of the plan, by tools and by agents, through one plan-store interface and `todo-item`, with the markdown file as one backend.
After item 344 every reader finds `todo.md` by its configured path, but each still parses the markdown itself, and agents still Read and grep the file. After this item agents list, show, add, edit, move and finish items only through `todo-item`, which gains list, edit and move commands. `rules/todo-items.md` loads today only when a session touches a `todo.md`, so it must load another way for a plan with no file. The interface carries what ranking needs: ID, release, blocked by, exempt, tagline, L, E, I, S, moves, detail, and an order the backend supplies (score order for markdown, Jira rank for Jira). One contract suite runs against every backend, on fakes only, per lilleman's call of 2026-11-12. R is above 5, so this item ships alone.
### 349. Convert markdown to ADF and back with `@larvit/adf-codec`'s portable calls, through a pinned Node shim run in a container with a memory limit.
lilleman's call, 2026-11-12: Jira Cloud descriptions and Confluence pages are ADF, and the loop keeps working in markdown. Write with `portableMarkdownToAdf` and read with `adfToPortableMarkdown`. The shim reads stdin and writes stdout, runs in a `node:<patch>` image with a pinned adf-codec version, and returns adf-codec's error codes as a Failure.
### 351. Read a project's plan from the child issues of a Jira Cloud epic its config names, ranked in Jira's backlog order.
lilleman's calls, 2026-11-12. One epic is one project's plan. An item is a child issue: its key is the item's ID, so branches, worktrees and commit and PR titles follow `CLAUDE.md`'s ticket rules (`ABC-123`, `ABC-123 - ...`). Summary is the tagline, description the detail (ADF, read through the adf-codec shim). Release is the Fix Version, or what item 359 configures. Blocked by is the "is blocked by" link. L, E, I, S, Moves and Exempt are issue entity properties, so no Jira admin is needed; the computed score also shows as a label such as `score-2.4`. Jira's backlog rank decides the order, since the team sets it; the loop's score doesn't break ties, and the loop doesn't reorder anything others filed. Auth: email plus API token from the project's env file.
### 353. Let a project's config take its goals and audience from a Jira epic's description, under the same two headings.
lilleman's call, 2026-11-12: this client keeps its goals and audience in the plan epic's description. The source is a config setting, never hard-coded; item 346 builds the file source this one adds to. Read through the adf-codec shim.
### 358. Let a project's config allow its runtime and tests on the host, and set the comprehension floor that marks code health very important.
lilleman's calls, 2026-11-12. Legacy products may have no container setup, so the config may allow host tooling; `CLAUDE.md`'s container rule and rule-hook then follow it. In 0.10.0 a decision in the project's records can put tooling on the host instead. The comprehension panel will score legacy code low, and today code health under 7.0 is very important. The config sets that floor, 7.0 in the configs of the projects under `~/code`; the signal stays that the score rises.
### 362. Send loop-claude's notices through the notifier a project's config names, with Mattermost as the first backend.
lilleman's call, 2026-11-12: each project names its notifier, server and recipient, so client notices can go to a client's workspace. Today `bin/mattermost-dm` reads one server from `~/.claude/.env` for every project.
### 363. Add a Slack backend to the notifier, sending each notice as a DM from a bot.
lilleman asked for a Slack driver on 2026-11-12, for clients on Slack. The bot token comes from the project's env file.
### 352. Write a project's plan to its Jira epic: file items as child issues, edit those the loop filed, score unscored issues with a comment, and move a finished item to the status its config names.
lilleman's calls, 2026-11-12. Without asking, the loop may create child issues in the epic, edit the summary, description and scores of issues it created, and comment on and transition any issue in the epic. It never edits the summary or description of an issue a person filed. An issue a teammate files without scores gets scored like any item, with a short comment saying it was scored and why; its scope stays as written. A finished item moves by the transition and to the status the config names, since each Jira workflow differs. An issue the loop files goes into Jira rank by its score among the issues the loop filed. A security finding is filed in the epic like any item, labelled `security` and ranked first, never with a secret's value. R is above 5, so this item ships alone.
### 355. Store a record on the Confluence Cloud page its config names, as markdown converted through the adf-codec shim, and interrupt the human when a teammate saved the page since the loop read it.
lilleman's calls, 2026-11-12: one page per record, whose ID the config names. The loop reads the page's ADF as portable markdown and writes markdown back through adf-codec. It writes with the version number it read; where Confluence reports a newer version, the loop asks the human, and never overwrites a teammate's edit on its own. Auth: email plus API token from the project's env file.
### 356. Hand reviewers the review standards from the folder a project's config names, in place of `~/.claude/skills/larv-review/standards/`.
lilleman's call, 2026-11-12: the folder replaces the default, as item 347's principles folder does. Readers: the reviewers under `agents/`, the judge's records, larv-review, and the repo check that keeps the standards present.
### 359. Tie a project's releases to what its config names: semver tags, Jira Fix Versions, sprints, or none.
lilleman's call, 2026-11-12: what a release is differs per client. With none, items name no release, and the architecture and principles passes the release review runs need another trigger, which this item settles and records. Readers: `rank_todo.py`'s release order, chunk-preflight's next release, loop-claude's "release ready to cut", larv-review's release review, `CLAUDE.md` → "Working loop". Cutting a release stays the human's.
### 361. Start the next item that doesn't touch a waiting pull request's files, in its own worktree and session, while that pull request waits for a human approval, and answer review comments on waiting pull requests as they arrive.
lilleman's call, 2026-11-12. Today a project has one session and one chunk at a time; this changes that core, so archive keys, loop-inbox, routing, the session lock and item 345's commits to the project repo all learn more than one open chunk. A chunk whose pull request gets a review comment resumes in its own session. R is above 5, so this item ships alone.
### 319. Start a fresh session with a brief of the chunk when a rung step changes the worker's model, in place of resuming the old session.
On SWE-bench Verified with Haiku 4.5 and Opus 4.7, handing the strong model the weak model's transcript solved 69.2% at $1.61 a task; starting the strong model fresh solved 79.2% at $0.90 (arXiv 2608.24358, The Handoff Tax). A model switch also rereads the whole context uncached (Claude Code prompt caching docs). A worker that steps from Sonnet 5.5 to Opus 5.5 by `--resume` is the raw handoff; an effort-only step within one model keeps the cache and isn't. The escalated subagent, which starts fresh with the trigger's evidence, already follows the pattern that won.
@@ -164,101 +256,9 @@ Add a `larv-review` subcommand that prints a round title's prefix, up to and inc
Left with the model: the summary after the `;`, and answering the findings.
### 343. Read every setting a project needs from one explicit config in the project's own git repo, `~/.claude-projects/<key>/`, found from the checkout's path, and fail naming each setting the config leaves out.
lilleman's calls, 2026-11-12. 0.11.0 is the bare minimum to run loop-claude on a client's legacy project: the plan, the records, the goals and audience and the principles move to paths the config names, and nothing else becomes pluggable yet. Each project has a folder `~/.claude-projects/<key>/`, where `<key>` is the project's name, and the folder is a git repo of its own. It holds the config, and for a client project the files the config points at. The config names the checkout it belongs to, so a tool run inside a checkout finds its config. A project is always a whole git checkout, never a subfolder.
No setting has a built-in default: a missing config, or a missing setting in one, fails and names what is missing. 0.11.0's settings are the paths of `todo.md`, `AGENTS.md`, `docs/decisions.md`, `docs/score-anchors.md`, `docs/comprehension-history.md`, `docs/reading-cost.md` and `CHANGELOG.md`, the goals-and-audience file (item 346), and the principles folder (item 347). A path may point into the checkout or into the project's folder. This chunk also writes a config for every project under `~/code` that names the files each one uses today, so each keeps running unchanged. The config's spelling is a syntax lilleman writes, so a panel seated as lilleman picks it before it ships. README → Install says how to set up a client project, on one sandbox VM per client.
### 347. Hand reviewers the technical principles from the folder a project's config names, in place of `~/.claude/principles/`.
lilleman's call, 2026-11-12: every project must set this folder. It replaces `~/.claude/principles/` and doesn't add to it, so the projects under `~/code` name `~/.claude/principles/`. The `Applies to` line works inside the named folder. Readers: larv-review's principles pass, the reviewers under `agents/`, and `CLAUDE.md` and the skills wherever they name `~/.claude/principles/`.
### 346. Read a project's goals and audience from the file its config names, in place of `README.md`'s `## Goals` and `## Audience`.
lilleman's call, 2026-11-12: for a client project, goals and audience live in a file of their own in `~/.claude-projects/<key>/`, with the same `## Goals` and `## Audience` headings and the same rules (`rules/goals-and-audience.md`), which must then load for that file too. A project reads only its own premises, never a parent project's. Every reader uses the configured file: loop-claude's prompts and judge, `rank_todo.py`'s goal weights and Moves check, larv-review's principles pass, the comprehension panel, and the places in `CLAUDE.md` that say `README.md`. A legacy project whose comprehension score would pull its plan towards refactoring can void code health in that file with the `**Void:**` line, as a README can today.
### 344. Read and write the plan, `AGENTS.md`, the records and the changelog at the paths a project's config names, in every tool, prompt, rule and skill.
Today about 40 modules, and `CLAUDE.md`, `rules/` and the skills, name `todo.md`, `AGENTS.md`, `docs/decisions.md`, `docs/score-anchors.md`, `docs/comprehension-history.md`, `docs/reading-cost.md` and `CHANGELOG.md` as files at the checkout's root. Among them are `rank_todo.py`, `todo-item`, `chunk-preflight`, loop-claude's routing, chunk items, judge records and prompts, larv-review, loop-inbox, `release_plan`, `subagent-rung` and `reviewer_context.principles_line`. After this item each of them uses the configured path. Agents get the path in the prompts and in the output of `todo-item` and `chunk-preflight`. Keep each file's name, so `rules/todo-items.md` and `rules/decisions.md` still load by their `paths:` patterns wherever the file lives.
lilleman's call, 2026-11-12: `AGENTS.md`'s path is configurable too. The tooling treats `AGENTS.md` as ours, since it lists the project principles and the decision titles, and a client repo may hold its own `AGENTS.md`, which stays theirs. Where the plan sits outside the checkout, loop-claude reads the chunk's items from that file, not from `todo.md` on the default branch. Tools that read `todo.md`'s git history in the checkout, `score-outcomes` and the routing figures, say they don't support a plan outside the checkout, and the loop goes on, until 0.13.0. R is above 5, so this item ships alone.
### 345. Commit each change a chunk makes to a plan or record outside the checkout into the project's `~/.claude-projects/<key>/` repo, titled like the chunk's own commits.
lilleman's call, 2026-11-12: each `~/.claude-projects/<key>/` is a git repo of its own, so plan and record edits keep a history. Those edits don't ride the chunk's branch: one loop runs per project, and it commits to the project repo's only branch. rule-hook's rank check, and the `Drop item` title rule, apply to these commits as they do in the checkout. Whether these repos are pushed anywhere for backup is a call for later: client data stays on the client's VM (goal 4).
### 348. Read each project's credentials from the env file its config names, fail naming each missing key, and keep every value out of prompts, transcripts, logs and the archive.
lilleman's call, 2026-11-12: credentials (Jira and Confluence API token, GitHub, Mattermost or Slack token) come from a per-project env file such as `~/.claude-projects/<key>.env`, outside the project's repo and named in the config. Which account each service uses is a config setting: a bot account or lilleman's own. A test shows a value never reaches a worker or judge prompt, a log line or the transcript archive, which guards goal 4. Today `~/.claude/.env` serves every project.
### 354. Read and write the decisions, score anchors, comprehension history, reading cost and changelog each through the backend its project's config names, with markdown files as one backend.
lilleman's call, 2026-11-12: each record has its own backend setting, so a client project can keep, say, decisions and changelog in Confluence and anchors and comprehension history as files. Which goes where is asked per client at setup. Every tool and agent that reads or writes a record goes through the interface: the judge's records, `rank_todo.py`'s decision check, `todo-item finish --changelog`, larv-review's scoring and places, `score-outcomes`, `rules/decisions.md`. The plan store's contract suite has a twin for records, on fakes. R is above 5, so this item ships alone.
### 364. Read a client repo's own `CLAUDE.md` and `AGENTS.md`, propose to its team each change that aligns them with this setup, and record each conflict left as a decision in the project's records.
lilleman's call, 2026-11-12: the loop may change every file in a client repo, but in agreement with the team, so a change to their instruction files goes to them as a proposal in a pull request. Where a conflict stays, the loop records which rule it follows and why as a decision in the project's decision record, so it is never asked twice. This is a working rule in `CLAUDE.md`.
### 360. Make `chunk-pr merge` wait for the human approvals a project's config requires, as part of the gate, before it merges.
lilleman's call, 2026-11-12: this client requires another human's approval on every pull request, so the gate covers it, and the loop still merges once the gate passes. Until this ships, lilleman merges such pull requests by hand and restarts the loop. On GitHub, read the approvals the branch protection asks for; this client's forge is GitHub.
### 357. Write prose in the language a project's config names, and title its commits, pull requests and branches by the templates it names.
lilleman's call, 2026-11-12: a client may want docs and Jira text in another language than English, and may use a commit, PR or branch format other than `CLAUDE.md`'s `ABC-123 - ...` and `ABC-123`. Everything reads the project's setting: `CLAUDE.md` → "Prose" and "Git, commits, and PRs", `chunk-preflight`'s branch and worktree names, `chunk-pr`'s titles, rule-hook's round-title checks and `score-outcomes`' `Drop item` titles. The prose reviewer judges prose in the configured language.
### 350. Route every read and write of the plan, by tools and by agents, through one plan-store interface and `todo-item`, with the markdown file as one backend.
After item 344 every reader finds `todo.md` by its configured path, but each still parses the markdown itself, and agents still Read and grep the file. After this item agents list, show, add, edit, move and finish items only through `todo-item`, which gains list, edit and move commands. `rules/todo-items.md` loads today only when a session touches a `todo.md`, so it must load another way for a plan with no file. The interface carries what ranking needs: ID, release, blocked by, exempt, tagline, L, E, I, S, moves, detail, and an order the backend supplies (score order for markdown, Jira rank for Jira). One contract suite runs against every backend, on fakes only, per lilleman's call of 2026-11-12. R is above 5, so this item ships alone.
### 349. Convert markdown to ADF and back with `@larvit/adf-codec`'s portable calls, through a pinned Node shim run in a container with a memory limit.
lilleman's call, 2026-11-12: Jira Cloud descriptions and Confluence pages are ADF, and the loop keeps working in markdown. Write with `portableMarkdownToAdf` and read with `adfToPortableMarkdown`. The shim reads stdin and writes stdout, runs in a `node:<patch>` image with a pinned adf-codec version, and returns adf-codec's error codes as a Failure.
### 351. Read a project's plan from the child issues of a Jira Cloud epic its config names, ranked in Jira's backlog order.
lilleman's calls, 2026-11-12. One epic is one project's plan. An item is a child issue: its key is the item's ID, so branches, worktrees and commit and PR titles follow `CLAUDE.md`'s ticket rules (`ABC-123`, `ABC-123 - ...`). Summary is the tagline, description the detail (ADF, read through the adf-codec shim). Release is the Fix Version, or what item 359 configures. Blocked by is the "is blocked by" link. L, E, I, S, Moves and Exempt are issue entity properties, so no Jira admin is needed; the computed score also shows as a label such as `score-2.4`. Jira's backlog rank decides the order, since the team sets it; the loop's score doesn't break ties, and the loop doesn't reorder anything others filed. Auth: email plus API token from the project's env file.
### 353. Let a project's config take its goals and audience from a Jira epic's description, under the same two headings.
lilleman's call, 2026-11-12: this client keeps its goals and audience in the plan epic's description. The source is a config setting, never hard-coded; item 346 builds the file source this one adds to. Read through the adf-codec shim.
### 358. Let a project's config allow its runtime and tests on the host, and set the comprehension floor that marks code health very important.
lilleman's calls, 2026-11-12. Legacy products may have no container setup, so the config may allow host tooling; `CLAUDE.md`'s container rule and rule-hook then follow it. In 0.11.0 a decision in the project's records can put tooling on the host instead. The comprehension panel will score legacy code low, and today code health under 7.0 is very important. The config sets that floor, 7.0 in the configs of the projects under `~/code`; the signal stays that the score rises.
### 362. Send loop-claude's notices through the notifier a project's config names, with Mattermost as the first backend.
lilleman's call, 2026-11-12: each project names its notifier, server and recipient, so client notices can go to a client's workspace. Today `bin/mattermost-dm` reads one server from `~/.claude/.env` for every project.
### 363. Add a Slack backend to the notifier, sending each notice as a DM from a bot.
lilleman asked for a Slack driver on 2026-11-12, for clients on Slack. The bot token comes from the project's env file.
### 352. Write a project's plan to its Jira epic: file items as child issues, edit those the loop filed, score unscored issues with a comment, and move a finished item to the status its config names.
lilleman's calls, 2026-11-12. Without asking, the loop may create child issues in the epic, edit the summary, description and scores of issues it created, and comment on and transition any issue in the epic. It never edits the summary or description of an issue a person filed. An issue a teammate files without scores gets scored like any item, with a short comment saying it was scored and why; its scope stays as written. A finished item moves by the transition and to the status the config names, since each Jira workflow differs. An issue the loop files goes into Jira rank by its score among the issues the loop filed. A security finding is filed in the epic like any item, labelled `security` and ranked first, never with a secret's value. R is above 5, so this item ships alone.
### 355. Store a record on the Confluence Cloud page its config names, as markdown converted through the adf-codec shim, and interrupt the human when a teammate saved the page since the loop read it.
lilleman's calls, 2026-11-12: one page per record, whose ID the config names. The loop reads the page's ADF as portable markdown and writes markdown back through adf-codec. It writes with the version number it read; where Confluence reports a newer version, the loop asks the human, and never overwrites a teammate's edit on its own. Auth: email plus API token from the project's env file.
### 356. Hand reviewers the review standards from the folder a project's config names, in place of `~/.claude/skills/larv-review/standards/`.
lilleman's call, 2026-11-12: the folder replaces the default, as item 347's principles folder does. Readers: the reviewers under `agents/`, the judge's records, larv-review, and the repo check that keeps the standards present.
### 359. Tie a project's releases to what its config names: semver tags, Jira Fix Versions, sprints, or none.
lilleman's call, 2026-11-12: what a release is differs per client. With none, items name no release, and the architecture and principles passes the release review runs need another trigger, which this item settles and records. Readers: `rank_todo.py`'s release order, chunk-preflight's next release, loop-claude's "release ready to cut", larv-review's release review, `CLAUDE.md` → "Working loop". Cutting a release stays the human's.
### 361. Start the next item that doesn't touch a waiting pull request's files, in its own worktree and session, while that pull request waits for a human approval, and answer review comments on waiting pull requests as they arrive.
lilleman's call, 2026-11-12. Today a project has one session and one chunk at a time; this changes that core, so archive keys, loop-inbox, routing, the session lock and item 345's commits to the project repo all learn more than one open chunk. A chunk whose pull request gets a review comment resumes in its own session. R is above 5, so this item ships alone.
### 365. Have `score-outcomes` read the finished and dropped items of a plan outside the checkout, from the project repo's history or from Jira's issue changelogs.
lilleman's call, 2026-11-12: 0.11.0 and 0.12.0 switch the history tools off for a plan outside the checkout, and 0.13.0 holds only their ports. In Jira a dropped item, which `Drop item` commit titles mark today, is an issue resolved by a resolution the config names.
lilleman's call, 2026-11-12: 0.10.0 and 0.11.0 switch the history tools off for a plan outside the checkout, and 0.13.0 holds only their ports. In Jira a dropped item, which `Drop item` commit titles mark today, is an issue resolved by a resolution the config names.
### 366. Count the chunks of a project whose plan sits outside the checkout in the routing figures, from the project repo's history or from Jira.