Lfg

CRITICAL: You MUST execute every step below IN ORDER. Do NOT skip any required step. Do NOT jump ahead to coding or implementation. The plan phase (step 1) MUST be completed and verified BEFORE any work begins. Violating this order produces bad output.

When invoking any skill referenced below, resolve its name against the available-skills list the host platform provides and use that exact entry. Some platforms list skills under a plugin namespace (e.g., compound-engineering:ce-plan); others list the bare name. Invoking a short-form guess that isn't in the list will fail — always match a listed entry verbatim before calling the Skill/Task tool.

Task Visibility

Before step 1, use the platform's task-tracking capability when available to publish a short stage-level view of the remaining pipeline. Derive it from the user-meaningful outcomes below rather than mirroring all ten steps or exposing internal gates. Before invoking a child skill, replace or clear LFG's view so only the child skill's task surface is visible; after it returns, recreate or refresh LFG's remaining pipeline work before invoking the next child. Add conditional work only when its gate fires. If no task-tracking capability is available, continue normally without simulating a task list in chat.

Artifact Root

This pipeline records the plan under <root>/plans/ and review residuals under <root>/residual-review-findings/. Resolve <root> when you first compose a <root>/ path (per the block below), never before you need it. A write to <root>/... and a read of <root>/solutions/ both count as composing a <root>/ path, so either one triggers resolution; only a run that touches no <root>/ path at all -- a scratch-only or no-repo flow -- skips it.

Resolve the CE artifact root <root> before composing any artifact path.

• Read docs_root from <repo-root>/.compound-engineering/config.local.yaml, then config.yaml; first non-empty value wins (<repo-root> = git rev-parse --show-toplevel). Unset -> <root> is docs, exactly as before.
• Validate a set value: a repo-relative directory whose real, symlink-resolved path stays inside the repo and is neither the repo root nor under .git/. Otherwise stop with an error naming docs_root and the value -- never fall back to docs.
• Use <root> as the sole artifact location: create it if absent, compose each path as <root>/<subdir> with this skill's own subdirectory, and never also read docs.

Per-stage routing carriers

Before step 1, interpret whether the invoking conversation expresses semantic intent to assign a pipeline stage — planning or implementation — to a specific model or harness. This is judgment, not keyword or prompt-token matching: an explicit instruction such as "plan with fable" or "use Codex for implementation" creates an assignment, while a plain mention of Codex, Composer, Fable, or another model/harness in feature content, quoted material, comparison text, or a filename does not. Two pipeline stages are routable, each with its own carrier:

• Planning routes to ce-plan as a plan_model:<alias> carrier — the plan-authoring model (model elevation), model-only. Example aliases: fable, opus. Planning has no cross-harness engine: an assignment that scopes a harness to planning ("plan with codex", "plan on cursor") is not supported — surface it as a routing-carrier blocker rather than encoding a harness name as plan_model:<harness>, which ce-plan cannot serve and would silently fall back to the session model. Only the implementation stage routes to a different harness.
• Implementation routes to ce-work as an implementation_engine object (grammar below) — the authoring harness/model.

Resolve each directive by scope:

1. Scoped directive — the instruction names the stage ("plan with fable", "codex for implementation", "plan fable, codex work"). Route it to that stage's carrier. Multiple scoped directives may resolve at once, each to its own stage.
2. Unscoped directive — a bare model/harness assignment with no stage named ("use fable", "with codex"). Bind it to the implementation stage only; never broaden an unscoped directive to planning or to every stage. Disclose the resolved binding in LFG's opening line before step 1 (e.g. "Routing implementation to Codex; planning stays on the session model.").
3. Unscoped and genuinely ambiguous, human present — when an unscoped directive could credibly belong to more than one stage and mis-binding would be materially costly, and the host is interactive (exposes a blocking-question tool and is not a disable-model-invocation/headless run), ask exactly one upfront question to bind the stage before step 1, then proceed hands-off. In a disable-model-invocation/headless run, never ask — apply the implementation default and disclose it. The default path is mandatory: LFG runs from schedulers, loops, and nested orchestrators with no user to answer, so an unresolved directive must always fall to the disclosed default rather than block.

Requirement strength is inferred from the whole instruction, not one word: "use Codex for implementation" is preference-strength (prefer); "only use Composer for implementation" is requirement-strength (require) because its meaning rejects native fallback.

Implementation carrier grammar. When implementation resolves to one candidate, retain one transient implementation_engine object with exactly these four fields:

• mode: prefer or require
• target: exactly one of codex, claude, grok, cursor, or composer — a harness name, never a model name
• model: the explicit model pin, otherwise null
• source: caller-visible provenance identifying the current LFG instruction

A directive that names a bare model with no harness (e.g. "use fable", "with opus") is a model pin, not a target: encode it as the harness that serves that model family with the alias in model — a Claude-family model (fable, opus, sonnet, haiku) is {"target":"claude","model":"<alias>"}. Never put a model name in target; if you cannot map the named model to one of the five harnesses, that is a routing-carrier blocker, not a null binding that silently drops the user's instruction.

When the implementation instruction instead names an ordered fallback list, do not truncate it to the scalar carrier — retain the whole ordered assignment as current-task implementation intent and pass no implementation_engine: object. At the CE Work seam, that still-active current-task assignment outranks config and is normalized/preflighted in order. This is stage-scoped context, not plan content; if the host cannot preserve that context across its skill invocation, stop with a routing-carrier blocker rather than silently dropping later candidates.

Sanitize product input. Remove every routing directive from the feature request that enters planning, keeping the request otherwise unchanged. Never pass the implementation_engine object or any removed directive to ce-plan, ce-doc-review, ce-code-review, the settled-decisions brief, or any planning or review product input — the carrier is stage-scoped routing authority, not product content or a settled product decision. The plan_model:<alias> carrier is the one exception: it is structured routing data handed to ce-plan alongside — never woven into — the sanitized request. Do not construct a carrier from standing configuration here: when no explicit binding exists for a stage, ce-work and ce-plan own resolution of still-applicable session/project intent and standing per-checkout configuration.

1. Invoke the ce-plan skill with the sanitized feature request prepared above (or the unchanged arguments you were invoked with when no routing directive was present). When a planning-stage directive resolved, prefix the invocation with its plan_model:<alias> carrier — structured routing data beside the request, never woven into it — so ce-plan's model elevation authors the plan on the chosen model even in pipeline mode.

   Before invoking, compose a settled-decisions brief from the invoking conversation and pass it with those arguments: direction (1-2 lines); settled decisions, each with four required fields — the decision, its provenance class (user-directed or user-approved), the rejected alternative, and a one-line reason; open areas; and a standing report-conflicts line. An entry whose rejected alternative cannot be stated demotes to a directive or open area. Scope topically — only decisions about the feature being shipped; when in doubt, demote (re-litigation is the safe floor; importing stale settlements is not). If the conversation contains no settled decisions, skip composition entirely and invoke ce-plan exactly as above — no empty-brief ceremony. The brief is transient: once ce-plan writes the plan, the plan's labeled KTDs are canonical.

   GATE: STOP. If ce-plan reported the task is non-software and cannot be processed in pipeline mode, stop the pipeline and inform the user that LFG requires software tasks. If ce-plan returned a blocked report containing settled-decision-invalidated, stop the pipeline and inform the user with the reason — do not retry. Otherwise, verify that the ce-plan workflow produced a plan file in <root>/plans/. If no plan file was created, invoke ce-plan again with those same arguments. The retry reuses the composed brief verbatim — never recompose it. Do NOT proceed to step 2 until a written plan exists. Record the plan file path — it will be passed to ce-work in step 2 and ce-code-review in step 4.

   Read the plan metadata before continuing. If the plan has artifact_contract: ce-unified-plan/v1, proceed only when it has artifact_readiness: implementation-ready and execution: code. Stop the pipeline for artifact_readiness: requirements-only, any unrecognized readiness value, execution: knowledge-work, approach-plan outputs, answer-seeking/universal outputs, or invalid progress-like readiness values. LFG never launches /goal directly; when goal-mode or dynamic workflows are appropriate, ce-work owns that implementation engine choice and must return control to LFG afterward.

2. Invoke the ce-work skill with mode:return-to-caller <plan-path-from-step-1> when no scalar transient carrier exists, including when a retained ordered current-task assignment is still active in context. When the scalar carrier exists, use the exact string-host form mode:return-to-caller implementation_engine:<compact-json> <plan-path-from-step-1>.

   If the scalar transient carrier exists, serialize its exact implementation_engine.{mode,target,model,source} data as compact JSON immediately after the implementation_engine: prefix (for example implementation_engine:{"mode":"prefer","target":"codex","model":null,"source":"lfg-current-turn"}). This is structured caller data in a portable string envelope, not part of the plan path or implementation prompt. Pass no empty carrier when it does not exist. ce-work then resolves a retained ordered current-task assignment when present, otherwise applicable session/project intent and standing per-checkout configuration. LFG is an automatic, headless caller: it never prompts to weaken a requirement-strength route.

   The optional implementation_run:<safe-id> carrier is recovery-only. Never include it on the initial step-2 call. On the one evidence-reconciliation recovery below, place it after the same engine carrier when one existed and before the unchanged plan path: mode:return-to-caller implementation_run:<safe-id> <plan-path-from-step-1> or mode:return-to-caller implementation_engine:<compact-json> implementation_run:<safe-id> <plan-path-from-step-1>. A safe id matches ^[A-Za-z0-9._-]{1,128}$ and contains at least one non-period character. Reject a malformed or duplicate run/engine carrier instead of launching work.

   GATE: STOP. Read the structured return before continuing. A status: blocked or status: failed return stops the pipeline. In particular, an unavailable require route must not prompt, fall back, or start native work. A completed prefer fallback may continue to step 3 exactly once after prominently disclosing its requested-versus-actual route/model and fallback_reason to the user; fallback is not a reason to invoke implementation again.

   For status: complete, verify that implementation work was performed - files were created or modified beyond the plan. Require the same plan path, changed files, U-IDs attempted/completed when present, verification results, behavior-change signal, and standalone_shipping_skipped: true. Also require the route-aware receipt fields implementation_engine_binding, requested_route, actual_route, requested_model, actual_model, fallback_reason, run_id, unit_receipts, plan_checkpoint, blockers, and recovery_path. These fields are required even when native execution makes some values null; together they carry binding provenance, requested-versus-actual identity, fallback, the durable run, per-unit process/integration/verification/commit state, checkpoint disclosure, blockers, and recovery. A resumed return must carry the same run_id; never treat resume as permission to start a new unit or a second LFG tail.

   When behavior_change: true, also require verification_evidence that names the relevant units/tasks, existing tests inspected, tests added/changed or used unchanged, red failure or characterization evidence when applicable, verification run, and any deliberate test exception. Do NOT decide the test strategy inside LFG; the evidence is ce-work's contract. Also read settled_decision_conflicts from the return: blocker-routed entries arrive as status: blocked and stop the pipeline; record any proceeded-and-flagged entries — they must reach step 6's durable residual record and step 8's PR-description context, since later review may not rediscover them.

   If behavior_change: true but verification_evidence is missing or too vague to tell how behavior was protected, invoke ce-work one more time in recovery mode. Reuse the same implementation_engine:<compact-json> carrier when one existed and keep the same plan path. With a safe non-null run_id, add implementation_run:<safe-id> with that exact value from the first return. When actual_route is native and run_id is null, repeat the original ce-work invocation once without an implementation_run: carrier; this preserves the pre-existing native idempotency/evidence-reconciliation path. A non-native return without a safe run id remains blocked instead of attempting discovery or a second implementation. Do not prompt the user and do not alter the plan path or engine carrier; this is evidence reconciliation, not a fresh dispatch. The recovery relies on ce-work's reconciliation path to inspect the already-implemented work, fill the missing evidence, and return without reimplementing. If the second return still lacks coherent verification evidence, stop as blocked and report the missing fields instead of continuing to simplify/review/ship.

3. Invoke the ce-simplify-code skill on the branch diff.

   This runs before review so the code-review in step 4 covers the simplified code. Skip this step when the change is docs-only (only markdown/docs paths changed) or trivial (roughly under 10 changed lines). Otherwise let ce-simplify-code resolve the branch-diff scope itself; it preserves behavior and runs the test suite. Pass the plan path from step 1 as structure-pin context, not as the simplification scope (the branch diff remains the scope), with a one-line constraint: session-settled:-labeled KTDs are structure pins the simplification must preserve (deliberate duplication stays duplicated).

   Do not commit in this step. ce-simplify-code leaves its changes in the working tree; step 4's review scopes the working tree (uncommitted changes included), and step 8's ce-commit-push-pr commits whatever remains. Committing here would sweep any still-uncommitted ce-work edits into a misleading refactor commit and could stall on a tree that never goes clean.

4. Invoke the ce-code-review skill with mode:agent plan:<plan-path-from-step-1>.

   Pass the plan file path from step 1 so ce-code-review can verify requirements completeness. Read the Actionable Findings summary the skill emits. Also read any findings stamped settled_conflict (each names the conflicting KTD). A stamped finding whose evidence is invalidating — the settled decision cannot work: infeasible, wrong-thing, or destructive — stops the pipeline as blocked, with the finding reported, before the shipping precondition. Stamped preference-grade findings proceed (they are report-only) but must flow into step 6's residual record.

   mode:agent is report-only by design — it surfaces findings but never edits the tree; LFG applies the eligible ones in step 5. When narrating progress to the user, frame this as "review found X → applied X in step 5," not as "code review did not auto-fix." A report-only review followed by an LFG-applied fix is the intended contract, not a gap.

Shipping precondition (steps 5–9). Run git remote once before the shipping steps. If it lists no remote (e.g. a sandbox/throwaway checkout that has git init but no origin), shipping is local-only: make every commit the steps below call for, but skip every push, PR create/edit, and CI-watch action — the pushes in steps 5 and 6, the push and PR creation in step 8, and step 9 in full. A missing remote is a terminal local-only state, not an error: never retry a push or hunt for a remote — make the local commits and proceed to step 10. Run steps 5–9 normally when a remote exists.

5. Apply and persist review fixes (REQUIRED after step 4, before residual handoff)

   Load references/review-followup.md and execute its apply step (mechanical apply + commit/push when changes exist). Do not proceed to the residual handoff, run browser tests, or output DONE while eligible review fixes remain only in the working tree uncommitted.

6. Autonomous residual handoff (only when step 4 reported one or more actionable downstream-resolver findings not applied in step 5; skip when it reported Actionable findings: none.)

   Do not prompt the user. This step embraces the autopilot contract: residuals must become durable before DONE, but the agent never stops to ask. Also run this step when step 4 emitted any settled_conflict-stamped findings, or when step 2's return carried proceeded-and-flagged settled_decision_conflicts entries — both sit outside the apply path, but they are the divergent class and must be made durable here.

   1. Load references/tracker-defer.md in non-interactive mode. Pass the residual actionable findings from step 4/5 (or the run artifact when the summary was truncated).
   2. Collect the structured return: { filed: [...], failed: [...], no_sink: [...] }.
   3. Compose a ## Residual Review Findings markdown section from the structured return (this goes into the committed record file in step 4, not the PR body):
      • For each item in filed: a bullet with severity, file:line, title, and a link to the tracker ticket URL.
      • For each item in failed: a bullet with severity, file:line, title, and the failure reason (e.g., Defer failed: gh returned 401 — tracker unavailable).
      • For each item in no_sink: a bullet with severity, file:line, and title inlined verbatim so the committed record file is the durable record.
      • For each settled_conflict-stamped finding from step 4: a bullet with severity, file:line, title, and the conflicting KTD the stamp names — included even though the finding is report-only.
      • For each proceeded-and-flagged settled_decision_conflicts entry from step 2: a bullet with the KTD, the evidence, and how it was routed.
   4. Durable record — never the PR body. Do NOT write a ## Residual Review Findings section into the PR description; it duplicates GitHub's own tracking and goes stale as items resolve. Review residuals have no GitHub thread of their own, so they are made durable by the tracker tickets filed in step 2 plus a committed record file — not a PR-body section and not a PR comment that duplicates the tickets. Create/replace <root>/residual-review-findings/<branch-or-head-sha>.md with the composed section (ticket links included) and the source run context. Stage only that file, commit docs(review): record residual review findings, and push when a remote is configured (per the shipping precondition): if an upstream exists, git push; else if a remote exists, resolve a writable one (prefer origin, otherwise the first configured remote) and git push --set-upstream <remote> HEAD; if there is no remote at all, the local commit is the durable sink.

   Do not output DONE until the residuals are durable (tracker tickets filed and/or the record file committed). Never block DONE on tracker filing failures once the record file exists. A push that fails when a remote exists is a stop-and-report; never retry a push, or block DONE, when no remote exists.

7. Invoke the ce-test-browser skill with mode:pipeline.

8. Invoke the ce-commit-push-pr skill with mode:pipeline branding:on. Thread the recorded plan path from step 1 into the invocation, along with any proceeded-and-flagged settled_decision_conflicts entries from step 2, so the PR body's settled-decisions provenance line and its proceed-under-flag clause can fire.

   This commits any remaining changes, pushes the branch, and opens a pull request — non-interactively, per the mode token. If it prints a New concepts: trailer after the PR URL, record the concept name(s) for step 10. Once the PR URL is known, back-fill it into any residual tickets filed in step 6 (the filed list) so each ticket links to the PR carrying the finding — best-effort, and never block DONE on a failed ticket update. If step 6 already opened a PR (check with gh pr view --json number,url,state 2>/dev/null), skip PR creation but still commit and push any uncommitted changes. Per the shipping precondition, when no remote is configured, do NOT invoke ce-commit-push-pr — its commit step pushes unconditionally (git push -u origin HEAD), so a literal invocation would still hit the impossible push. Instead commit any remaining changes locally yourself (git add -A && git commit) and skip the push and PR creation entirely.

9. Watch the PR to CI-decided via ce-babysit-pr (only when an open PR exists for the current branch)

   Detect the PR; if none exists or gh is unavailable, skip this step entirely and proceed to step 10.

   gh pr view --json number,url,state
   

   Invoke ce-babysit-pr mode:pipeline <pr-url>. It runs the bounded pipeline loop: watches CI, repairs real (convergent) failures via ce-debug mode:pipeline — never weakening, skipping, or mocking an assertion — resolves any review comments that arrived via ce-resolve-pr-feedback mode:pipeline, and stops when CI is decided or its budget (default 3 fix rounds) is hit. This replaces LFG's former hand-rolled CI loop; do not reimplement CI-watching here. Invoke it unconditionally whenever an open PR exists — a run whose CI looks likely-clean is not a reason to skip babysit and poll gh pr checks yourself. Green CI at one instant is not this step's goal: babysit also resolves review comments across the PR's life, so a passing check while advisory checks (e.g. Bugbot) are still pending or comments are unhandled is not "done" and never substitutes for the invocation.

   Collect its structured result ({ status, fixes_applied, residuals }). It surfaces unfixable CI as a run-report comment on the PR and returns residuals — do NOT write a ## CI Failures Unresolved PR-body section. A needs-human residual (a fix that would need a product/design decision) is deferred, not applied — that is the autopilot contract, unchanged. Do not block DONE once babysit has surfaced residuals.

10. Output <promise>DONE</promise> when complete

    For the two user-runnable handoffs below, default to /ce-explain <name> / /ce-babysit-pr <pr-url>. Use $ce-explain <name> / $ce-babysit-pr <pr-url> only when the active host is Codex or explicitly documents dollar-prefixed skill invocation. Render only the invocation as inline code and output one form only.

    If step 8 recorded a New concepts: trailer, first echo one line per concept: New concept introduced: <name> — run <rendered ce-explain invocation> to go deeper.

    If an open PR exists, add one line pointing the user to the interactive watch-to-merge (pipeline mode stopped at "CI decided," not "merged"): PR is moving — run <rendered ce-babysit-pr invocation> to watch it through review to merge.

    Before the DONE promise, inspect the canonical plan from step 1 for the semantic role work-relationships. Load references/next-work-handoff.md when that role exists, or when an older unmarked Product Contract appears to name the area this plan owns plus future separately planned areas and their relationships; the reference owns the cautious legacy semantic fallback, candidate selection, and opt-in offer contract. Do not match an exact visible heading, treat ordinary non-goals as future work, or invoke ce-handoff before the user explicitly accepts the offer. If neither semantic signal exists, do not load the reference and make no next-work offer.

    Then output the DONE promise.

Start with step 1 now. Remember: plan FIRST, then work. Never skip the plan.

Source: EveryInc/compound-engineering-plugin → skills/lfg/SKILL.md