feat(lfg): route each request to the skill that owns it before shipping - #1702
Conversation
lfg now reads the request before it plans. A reproducible defect goes to ce-debug in a new return-to-caller mode that fixes on a feature branch and commits without pushing; an unsettled product shape goes to ce-brainstorm in a new return-to-caller mode when a human is present; a judgment goes to ce-pov first and only Adopt or Trial continues; a request whose result is not a code change runs that sibling and ends there. A plan ce-plan wrote in the same session goes straight to ce-work, and lfg never executes a plan file it merely found on disk. The body is restated as an outcome spine with conditions instead of ten numbered steps with STOP markers, ce-compound runs before the PR opens so the learning lands in the watched head, merging is the user's unless granted, and the internal word "tail" is gone from the skill and its references (shipping-tail.md is now shipping.md). Claude-Session: https://claude.ai/code/session_0199q8oLAGsFn6iYU3gu1awy
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: 180b1186c8
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".
… attended run reach ce-prototype Address PR review feedback (#1702) - ce-debug's return-to-caller result now carries pre_fix_scope (already-dirty files and pre-existing unpushed commits); lfg passes the dirty files as an exclude list to ce-commit-push-pr, scopes simplify and review to the fix, and keeps the fix local when the branch carries commits the user did not offer. The no-remote path commits files by name instead of git add -A. - ce-prototype stops on "no human present" rather than on the caller being lfg, and lfg's prototype route requires a human present. Claude-Session: https://claude.ai/code/session_0199q8oLAGsFn6iYU3gu1awy
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 3435b59148
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".
…eans an open PR contains it Address PR review feedback (#1702), round 2 - Route 3 applies only when the request asks for a change that hangs on an unsettled judgment; a judgment with nothing to build is route 5 and ends with ce-pov's verdict instead of rolling into planning. - The defect route's "ship only what the user offered" block is restated as its condition: prior branch work is offered only when an open PR already contains it. ce-debug returns prior_commits_ahead_of_base in place of an unpushed count, since a backup push does not offer anything. Claude-Session: https://claude.ai/code/session_0199q8oLAGsFn6iYU3gu1awy
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 35358b19ca
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".
Address PR review feedback (#1702), round 3: route 1 now requires that the request asks for the plan to be carried out or continued; a request that only mentions a plan while asking for something else is routed by what it asks for. Claude-Session: https://claude.ai/code/session_0199q8oLAGsFn6iYU3gu1awy
…cts, not verdicts Three review rounds each found a new edge in the defect route's shipping gate (containment proof, push ordering before the review push, where residuals go when shipping is withheld). Each was a symptom of a mechanism-shaped rule. The block now states the goal once, before any push in the run: publish only what the user offered, and hold when that cannot be established. ce-debug's pre_fix_scope records what it observed (pre-fix HEAD, dirty files, commits beyond base) and carries no verdict; the caller decides from it. Claude-Session: https://claude.ai/code/session_0199q8oLAGsFn6iYU3gu1awy
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f7a5e43097
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".
- ce-debug return-to-caller stops before editing a file that already carries the user's edits, instead of editing and then refusing to commit. - Defect-route review: ce-code-review has no fix-only scope, so step 5 applies findings only to the fix and this run's own edits; the rest are residuals. - Judgment route: what continues is a verdict that supports the change, not only Adopt/Trial, matching intake route 3 for approach-set positions. - Plan brief: skip composition only when there is nothing to carry, so ce-pov evidence reaches ce-plan even with no settled decisions. Claude-Session: https://claude.ai/code/session_0199q8oLAGsFn6iYU3gu1awy
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f58f4677d8
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".
…ototype handoffs - Shipping precondition sits before the first push (step 5), not after step 7. - An existing PR still receives the description context: pipeline mode leaves the body alone, so the default path re-invokes description-update mode. - A brainstorm brief carries every decision that meets the settlement test; intake passes them into the settled-decisions brief. - A prototype-then-build request continues into the plan route with the returned choices as user-directed decisions instead of ending the run. Claude-Session: https://claude.ai/code/session_0199q8oLAGsFn6iYU3gu1awy
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: dc48c477b4
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".
…eaches ce-plan Reproduction is ce-debug's question, so intake no longer requires a reproducible symptom to take the defect route. The brainstorm return's grounding_path is passed to ce-plan on both continuations. Claude-Session: https://claude.ai/code/session_0199q8oLAGsFn6iYU3gu1awy
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: e95f07517e
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".
…o can answer - A diagnosis-only request is route 5, not the fix route. - Human present means a person can answer in this session, whatever the host's question mechanism; the tie-break rule matches the concrete-report condition. - ce-debug return-to-caller verifies per fix.md, which already allows the reproduction check or characterization when a red test cannot run. Claude-Session: https://claude.ai/code/session_0199q8oLAGsFn6iYU3gu1awy
1 participant
Summary
lfgnow reads the request before it plans, and takes it to the end state its shape calls for through the skill that owns that work. A bug report goes toce-debug, an unsettled product idea goes toce-brainstormwhen you are present, a judgment goes toce-povfirst, an explain or prototype request runs that skill and ends there, and a plance-planwrote earlier in the session goes straight toce-workinstead of being planned twice. Before this, every/lfginvocation was planned as a feature, whatever it asked for.The body is restated as an outcome spine with conditions rather than ten numbered steps with STOP markers, per the repo's skill-authoring standard. Every gate the old body carried survives as a condition, and every test pin on the shipping seams is unchanged.
What changes for someone using it
/lfg fix this bug ESP-1234root-causes and fixes on a feature branch, then reviews and ships. It never writes a plan for a bug. It publishes only what you offered: on a branch carrying work no open PR holds, the fix is committed locally and nothing is pushed./lfg diagnose #123returns the diagnosis and stops./lfg explain the architecture of Xor/lfg create a prototype exploring 3 conceptsrunsce-explainorce-prototypeand ends with that result. No branch, no PR. A request to prototype and then build the winner continues into planning with the chosen decisions./ce-plan …then bare/lfghands the plan directly toce-work. A requirements-only artifact from/ce-brainstorm, or a path to an older plan, still goes throughce-planso drift and settled-decision conflicts are checked.lfgnever executes a plan file it found on disk. The plan is one the session identifies.ce-compoundruns before the PR opens, so a captured learning is in the head CI watches rather than pushed after "CI decided".lfgnever merges.Design decisions
mode:pipeline.ce-debug's pipeline mode pushes its own commit and skips review because the babysitter owns the PR; underlfgthe caller owns push and review, so the new mode commits on a feature branch and returnsfixed | diagnosed-no-fix | needs-human | blockedwith root cause and verification evidence.ce-brainstorm's new mode skips its handoff menu, which would otherwise re-invokelfg, and returns the artifact path or chat brief.ce-povverdict is cited evidence, never a settled decision. The settled-decisions design allows only user-directed and user-approved provenance and says an agent never self-settles its own recommendation. A verdict that supports the change continues (Adopt or Trial on an adoption question, a supporting position on an approach set); anything else stops the run with the verdict.ce-planin pipeline mode, which records the fork as an assumption in the plan. That isce-plan's existing contract; stopping instead would be a change to that skill.lfgstays model-invocable. Flagging itdisable-model-invocation(orallow_implicit_invocation: falseon Codex) would break the brainstorm handoff on both hosts, the reason PR fix(lfg): allow agent invocation for the autonomous brainstorm handoff #1116 removed the flag. The description is the activation control, shortened from 537 to 350 characters.ce-planreported writing this run" guards against adopting a discovered file and keeps that job on the plan route; a same-session plan is not a discovered file and is verified byce-work's content check instead.Validation
bun run test(4,103 pass),release:validate, andplugin:validateon the committed state.intake.mdbefore choosing, and never opened the plans directory. Pre-change baselines for the bug and bare cases showed the old plan-first behavior.ghshim, whole plugin loaded) on Claude and Codex: bug by reference, same-session plan then bare/lfg, explain, judgment, and a divergent bug (a test asserting a documented contract) which returnedneeds-humanwith nothing edited on both hosts. The interactive brainstorm seam ran end to end on Claude through an Orca terminal with the author answering as a stand-in.ce-work,ce-debug, andce-plan; explicit hands-off requests reachedlfg.ghshim cannot serve. Codex applies thece-debugstructured return from context rather than printing it, so the field-by-field return contract is proven on Claude only.The full evidence trail and the proposal's revision history are in
docs/brainstorms/2026-09-13-lfg-front-half-proposal.md.Review rounds
Seven Codex rounds. Rounds 1 to 3 asked for containment proof, push ordering, and a residual sink on the defect route's shipping gate; those were parked as one decision and answered by restating the block as a single goal (publish only what the user offered, decided before any push) with
ce-debugreturning observed facts and no verdict. Rounds 4 to 7 were handoff gaps, each fixed by one condition: stop before editing a file the user already changed, no fix-only review scope so step 5 applies findings only to the fix, brainstorm briefs keep settled decisions and grounding, an existing PR still receives the description context, a diagnosis-only request is not the fix route, attendance means a person who can answer, and verification defers toce-debug's own fallback.Related
Related: #879. That PR proposed a
ce-compoundstep inlfgplus dogfood, feedback-as-input, and reproduce-before-fix; this change lands the compound step and the reproduce-before-fix outcome throughce-debug, and leaves dogfood and feedback-as-input out.Follow-ups named in the recorded proposal: a single-PR merge grant has no carrier to
ce-babysit-pryet; thece-plance-explaincondition is untouched pending its own eval;ce-work's prose still calls its shipping sequence a "tail".Security Disclosure
No security-relevant changes. The new
ce-debugmode commits only fix-owned files on a feature branch and never pushes; no shell recipes were added to any skill.Agent Disclosure
Claude Code · claude-fable-5-1https://claude.ai/code/session_0199q8oLAGsFn6iYU3gu1awy