M79: plans as files — save the approved plan, implement it in a fresh conversation - #53
Open
RandyNorthrup wants to merge 11 commits into
Open
RandyNorthrup wants to merge 11 commits into
RandyNorthrup wants to merge 11 commits into
Conversation
In Plan mode the latest reply gets Save plan and Implement in a fresh conversation; pressing either is the approval. Neither backend marks a plan on the wire: a live Muse Code 1.4.0 Plan-mode capture (19 attempts) showed the plan as an ordinary agentMessage wrapped in its bundled plan skill's handoff, and MSP has no todo-set command. - Save writes the plan byte for byte to .agents/plans/YYYY-MM-DD-<slug>.md, the location Muse Code's own plan skill names (D13), with a numeric suffix on a taken name. It is published by a hard link from a hidden stage (createFileExclusively, now shared with memory), confined to the workspace's own .agents/plans, asks first (.agents is protected, D24), and is refused in Restricted Mode. A Muse Code plan reply's two handoff lines are left out. - Implement starts a new conversation from a brief (ConversationBrief, startFromBrief, for M74's /handoff): the plan file as named text, a MODEL_TEXT note, nothing else of the old conversation, the starting permission mode. On the Model API the plan's steps become the todo list first (AgentSession.setTodos); on Muse Code the note asks the model to take them as its list. - Plans... in the palette lists the saved plans to open or implement. - 25 strings in fifteen tables, harness scenarios plan, plan-brief and plan-narrow, 17 red drills, a live Model API case (7 requests); npm run quality exits 0. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…iles Brings in D49's settled rules (2407109), among them "Untrusted content": a conversation built on untrusted content starts in a mode that asks, whatever initialPermissionMode says. The M79 review fixes build on it. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Merges origin/plan/d49-competitive-program (2407109) for D49's settled "Untrusted content" rule, then fixes every finding of the three class reviews of 059ea2a in one pass. - A plan picked from Plans... is untrusted content: it starts in Manual (Plan when that is the starting mode), whatever initialPermissionMode says, and its note tells the model nobody confirmed who wrote it. Only a reply saved from a Plan-mode turn here is "approved"; even that never starts in Bypass in a remote window. - Save and Implement read the reply back on every press, resume the conversation after a restart (resumeTarget), and say why when they do nothing (another conversation, history not served, too large, a busy action, saved but not started). A reply counts only when its turn was sent in Plan mode, not steered, and finished with Plan never left. - The same plan saved already is found by its bytes, not written twice. A plan with HTML the panel hides is saved with a warning, not started. - createFileExclusively checks the folder again after mkdir and before the link, for plans and memory notes (a junction swap after the check is refused); a file where the folder should be says so; a held stage is removed again and stale stages swept. Plan files keep the plan limit even when they start like a PDF; names refuse control and format characters; cuts keep whole characters. - The model reads English MODEL_TEXT and, on the Model API, the steps its todo list was set to; a failed brief takes its todo list and chip back. The log names a plan by its date and a hash. - 9 strings in fifteen tables, the plan-brief harness as the Model API, 75 red drills (SHA-256 restored), live case19 through the panel's controller (10 requests, about $0.0008); npm run quality exits 0. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Owner
Author
|
@codex review |
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 335a533fff
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
… a refused mode change Codex on PR #53: - P1: hasHiddenMarkup found fences by a line prefix, so a backtick fence whose info string holds a backtick hid the lines after it while the panel rendered them, an HTML comment included. The plan is now parsed with the panel's own grammar (mdast-util-from-markdown + micromark-extension-gfm + mdast-util-gfm, the versions react-markdown and remark-gfm resolve to) and every html node is flagged. Swept: the title and the steps come from the same tree; link and picture titles and definitions nothing refers to, which the panel never shows, are flagged too, and the notices say so in all 15 tables. A jsdom test checks the flags against MarkdownView's own render. - P2: pending plan turns were cleared before setApprovalMode succeeded; a refusal now puts them back (a turn that finished meanwhile becomes a plan turn). Swept: updateEffort no longer shows an effort the session refused, unless the new model no longer serves the old tier. Pinned in PLAN.md D3; +114.5 KiB host bundle (563.2 of 600 KiB); character-entities added to the notices. 19 red drills, SHA-256 restored. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The panel's Markdown parser (114.5 KiB) no longer rides in the activation bundle, following M57's pattern: src/core/plans/planMarkdown.ts is built from src/host/planMarkdownEntry.ts into dist/planMarkdown.js (114.7 KiB, budget 150 KiB) and required by planMarkdownLoader on the first Save plan, Implement or Plans...; dist/extension.js is 449.7 KiB again. - No fallback: a reader that cannot load refuses the plan action with planMarkdownUnavailable (15 tables) and is tried again next time; the hidden-text check is never skipped. - The reader imports no value of shared/constants (it would bring the English table): it lists items, planDocument.ts cuts and caps them. - check-bundle-split fails when dist/extension.js carries the reader, its entry or any file of the parser's packages; budget, host-globals check, .vscodeignore, notices, knip, dpdm and an integration test cover the new bundle. requireFile/forgetFile moved to host/lazyBundle.ts. - Tests: the real built bundle, a missing and a wrong one, and the controller refusing Save, Implement and Plans...; the jsdom parity test stays. 26 red drills, SHA-256 restored (one the split gate itself). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Owner
Author
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 6729c3f07a
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Codex's third round on PR #53: a link's destination was sent in the approved brief but not shown (P1), and the log took the first ten characters of any plan name (P2). Per the owner's rule, a redesign: - One rewritten tree for both sides. showPlanParts (shared/planView.ts) turns every part of a parsed plan into rendered text: a link's destination after its text as <https://...>, a picture's alt text and source, titles, definitions and footnotes as paragraphs, a code fence's whole info string as its label. The panel renders the reply the plan actions sit under through it (MarkdownView isPlan, a remark plugin); the brief is the same tree written back as Markdown (briefText, mdast-util-to-markdown 2.1.2), for a reply and for a file. - The hidden-markup predicate is gone; only raw HTML, which the panel never renders, is found (hasRawHtml) and refused as before. The notices say HTML again in all 15 tables. - A jsdom test renders 13 tricky plans in MarkdownView and checks that every text, code and code-info leaf of the brief appears in order in the DOM text, and that the brief holds no link, picture, definition or footnote a view could show only part of. - planLogName: YYYY-MM-DD-#hash.md only for a verified real day, else #hash.md; the memory note's stage warning no longer names the note. - A plan file must be UTF-8 text without control characters. - 17 red drills, SHA-256 restored. dist/planMarkdown.js 139.0 KiB (budget 150), dist/extension.js 464.8 KiB after merging main. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Grok Build on 6fe3e07: - P1 fence info string: rejected on inspection. hast-util-to-jsx-runtime 2.3.6 passes a data-* property under its attribute name (lib/index.js:617), so the panel reads data-info and labels the block with the whole info string (rendered in jsdom). The review's case joins the parity corpus and a test asserts the label; reading dataInfo turns both red. - P1 control and format characters: a direction override or zero-width character is painted otherwise than the model reads it. hasUnshownCharacters (Cc but tab/LF/CR, DEL and C1 included; Cf) refuses Save and Implement of a reply and Implement of a plan file, with planUnshownCharacters (15 tables). - P2 PLAN.md D3: the measured sizes (planMarkdown.js 139.0 KiB, extension.js 464.8 KiB). 6 red drills, SHA-256 restored. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…uct call for the owner PR #53's two open Codex threads (on 6729c3f) were fixed by 6fe3e07: link destinations are shown in the plan view and briefed from the same tree, and planLogName keeps a prefix only for a verified day. This pins the thread's exact name (customer-secret.md) in the log test and states in the cert and README that plans with any control or format character (emoji joiners, LRM/RLM included) are refused, a product call to make. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Owner
Author
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 01f7f7639b
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
…gs no path Codex on 01f7f76: - P2 a view/gap reload rebuilt user cards without isPlanTurn, so Save plan and Implement vanished while the host still held the plan turn. historyLoaded now carries planTurnIds (finished or still running Plan-mode turns of those items) and the panel's replay marks their cards. isPlanTurn is the only per-card flag M79 added. - P2 Plans... Open called openFile outside the plan failure path, so a plan deleted after the pick reached the general handler, which logs the error detail (its path). It now fails through planFailed: planOpenFailed (15 tables) and the reason in the panel, only the error kind in the log. Every plan operation's failure path now does the same. 6 red drills, SHA-256 restored. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Owner
Author
|
@codex review |
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
M79 (PLAN D49): a plan the user approved survives and can drive a clean run, on both backends.
.agents/plans/YYYY-MM-DD-<slug>.md, Muse Code's own convention (its bundledplanskill in 1.4.0), recorded in D13..agentsis a protected path, so saving asks first. Saving never replaces a file (hidden stage + hard link), and the folder is re-checked against link/junction swaps.startFromBriefpath that M74's/handoffwill reuse. On the Model API the plan's steps seed the todo list; on Muse Code the brief asks the agent to list them itself.Trust
Evidence
docs/certification/m79.md). 75 red drills.npm run qualityexited 0 (run alone). Muse Code Implement was not run live, as the cert says.🤖 Generated with Claude Code