Repository navigation
feat(m8): assemble operator procedures from named fences, with lints - #214
Conversation
Live sessions have stopped on composition defects, never on the code under test: a blanket empty-stderr rule against a CLI summary, a block hardcoding the previous generation's paths, quoting at a transport boundary. Each was caused by hand-copying blocks into step scripts. scripts/m8_procedure_package.py reads one TOML Composition block in a reviewed procedure document. It builds exact LF step files, their declared stdout and a SHA-256 manifest from named fences in hash-pinned frozen sources. It refuses, naming the step: - unresolvable sources, headings or fence indexes; - a lane CLI call outside lane_capture, which would have caught the g11 S4 stop; - undeclared or historical generation labels, which would have caught the H7 path gap; - unreviewed placeholders outside --template; - CRLF documents. It never runs a step. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01S7VMTaJ9XePzc8fjmBBFfs
…nces Review of 5488dd8 (Codex) found, all fixed here: - The line-based lane lint was bypassable (`lane_capture x true; "${LANE[@]}" ...`, a command substitution, any backgrounded call that discards its output) and also refused harmless forms. It is replaced by an allowlist with no shell parsing. Only the reviewed blocks may name the lane array: #210 L1-L3, plus OC Setup, S1 and S6a. Literals never may, and splitting the name with a backslash-newline does not hide it. - Permissive TOML shapes changed meaning: `historical = "false"` was truthy, `stdout = "ok"` became two lines, empty parts built an empty script, and a mixed literal/selection part ignored its selector. The schema is now strict. - LF was checked only on input. Escaped CR in literals or stdout is now refused. - Indented fences were invisible, four-backtick examples exposed their inner blocks, and a closer with a suffix closed the block. A CommonMark-style scanner fixes all three, and headings inside fences are ignored. - The UNPREPARED refusal sentinel was a false placeholder hit. Only assignments (`=UNREVIEWED`, `=UNPREPARED`) count now. - `check` is strict unless --template. - Selected fences must be bash. Declared stdout is bounded to 65,536 bytes. Output is built in a sibling directory and renamed into place. Labels stay a spelling check, documented as such, not proof of fresh paths. Refs #186 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01S7VMTaJ9XePzc8fjmBBFfs
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. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 97843e8d12
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".
…im OUT Connector review of 97843e8 found these, all fixed: - Quoted placeholder assignments (RUN_ID="UNREVIEWED") passed the check. - Malformed sources (sources = [], a non-string path, a source named self) raised a traceback instead of a refusal. - An empty literal part was silently accepted. - On POSIX the final rename could replace an empty OUT that appeared meanwhile. OUT is now claimed with an exclusive mkdir, files are created with x mode, and MANIFEST.sha256 is written last, so a package without a manifest is incomplete. Refs #186 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01S7VMTaJ9XePzc8fjmBBFfs
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 148aacef14
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".
…labels Connector review of 148aace found these, all fixed: - The allowlist matched by basename, so a pinned copy at fake/<name> passed. Only a bare sibling path of the procedure document can now be an allowlisted reviewed source. - steps = 1 and a non-string source selector raised tracebacks. They are now refused. - A label split by a line continuation (q\<newline>11) escaped the label check. Labels are now matched after joining continued lines. Refs #186 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01S7VMTaJ9XePzc8fjmBBFfs
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d39f8f144a
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".
…uble labels Connector review of d39f8f1 found these, all fixed: - Placeholders were checked on unjoined text, so RUN_ID=UNRE\<newline>VIEWED slipped through. - An empty selected bash fence produced a zero-byte step. - A label declared both current and historical was silently current. Refs #186 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01S7VMTaJ9XePzc8fjmBBFfs
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 13d0f09d65
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".
…s exact Connector review of 13d0f09 found these, both fixed: - Shell escaping and concatenation ($'X', UNRE\VIEWED, "UNRE""VIEWED") still hid a placeholder. Quotes, backslashes and $ are now removed before matching, and UNREVIEWED anywhere, or UNPREPARED anywhere but the UNPREPARED.bundle sentinel, is refused. - Literal parts lost extra trailing newlines. Their bytes are now kept exactly, and only a missing final newline is supplied. Refs #186 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01S7VMTaJ9XePzc8fjmBBFfs
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 8df8e21093
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".
Connector review of 8df8e21: - UNPREPARED.bundle.old passed as the sentinel. Only the complete UNPREPARED.bundle token is exempt now. Fixed. - q'1'1 hid a stale label. Labels and placeholders now share one shell-normalized text, with continuations joined and quotes, backslashes and $ removed. Fixed. - ANSI-C hex escapes are recorded as a design choice, not handled. The lints catch honest composition mistakes in reviewed documents; deliberate obfuscation is the review's job, not a shell interpreter's. Refs #186 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01S7VMTaJ9XePzc8fjmBBFfs
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 2961af00f3
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".
Refs #186
Why
Every recent live M8 session stopped on a defect at the procedure's composition or transport boundary, never on the code under test:
cp -nwarning.All of them came from step scripts copied together by hand.
Change
scripts/m8_procedure_package.pytakes a reviewed procedure document's single TOMLCompositionblock. It builds:MANIFEST.sha256, with the manifest's own SHA-256 printed.Steps are built only from named fences in hash-pinned frozen sources and the document itself. Nothing is copied by hand. It refuses, naming the step:
LANE) unless it comes from an allowlisted, reviewed lane-boundary block: Reconcile operator procedure output checks with real lane CLI summaries #210 L1–L3, OC Setup, S1 and S6a. Literals never may. The real frozen OC S4 and ER R1 bytes are refused;bash. Fences are parsed CommonMark-style, so indented fences, longer fences and exact closers are handled, and headings inside fences are ignored;=UNREVIEWEDor=UNPREPAREDplaceholder, outside--template;It never runs anything. It's the first half of the end-to-end rehearsal; the g12 procedure, its Composition and a fake-host rehearsal of the built steps follow in the next PR.
Review
One Codex pass (gpt-6.1-sol, high) at
5488dd8found introduced defects. All are fixed in97843e8:lane_capture x true; "${LANE[@]}", command substitution, any backgrounded call that discards its output) and also refused harmless forms. It is replaced by a source allowlist that needs no shell parsing.checkwas not strict.One finding was found independently and also fixed: headings inside fences.
Accepted as a design limit and documented: labels are a spelling check.
Evidence
tests/test_m8_procedure_package.py: 28 passed locally.The design was checked with Fable before implementation. It adopted:
🤖 Generated with Claude Code
https://claude.ai/code/session_01S7VMTaJ9XePzc8fjmBBFfs