Worker bridged implementer path (#431 follow-up)

[DanielKillenberger]

Worker bridged implementer path (#431 follow-up)

Step 3c of the work skill routes implementation to the implementer tier, but the only thing it ever spawns is the in-host worker subagent, and the worker's own phases never consult that tier. When a project's routing block pins the implementer to a model the harness reaches only through a CLI bridge, the two documents disagree by omission.

Spec: fn-245-worker-bridged-implementer-path-431
Branch: fn-245-worker-bridged-implementer-path-431 → main
Tasks: 1 completed
R-ID coverage: 7/7 evidenced

Follow-up to #436, which closed #431.

TL;DR

• When the implementer tier is pinned to a model this harness reaches only by CLI, the in-host worker now runs the bridge itself; the conductor never bridges.
• The worker anchors and records the base commit as today, skips its own investigation and any scout fan-out, and hands the child a pointer prompt plus the usage guide's timebox-free long-task brief; the worker's own TIMEBOX is never copied into that brief.
• On return the worker commits anything the child left uncommitted, reviews the child's commit range, runs the gates on it, and continues with review dispatch and flowctl done exactly as before.
• The minted-task subagent license narrows on this path: parallel bridge calls on disjoint surfaces, each child in its own isolated workspace and branch; research and scouting are the child's job.
• The Codex mirror is regenerated, and its step 3c now carries the implementer-tier paragraph it had been dropping since Orchestration by intent: named tiers, reach pages, zero shipped slugs #352.

Not in this PR (by design)

• No flowctl code, hook, guard, or config key; the path is prose, consistent with Bridged implementer may commit checkpoints; timebox-free brief for long bridged tasks (#431) #436.
• The bridge safety rule and the long-task brief in the usage guide are not changed; the worker references them.
• No change to the conductor's dispatch template fields, the parallel-wave and host-deferred contracts, the review backends, or flowctl done.
• No change to the reach pages or the orchestration guide beyond what a reader needs to find the worker path.
• No site (flow-next.dev) changes in this PR.

The change, top to bottom

When the implementer tier resolves to a model the harness reaches only by CLI bridge, the in-host worker now runs the bridge itself: it anchors and persists the base as today, skips its own investigation and any scout fan-out, hands the child a pointer prompt plus the usage guide's timebox-free brief, and on return reviews the child's commit range and runs the gates before review dispatch and flowctl done. The conductor never bridges, and the minted-task subagent license narrows on this path. Prose only; the Codex mirror is regenerated.

Proof	Value	Sources
Artifact	fn-245-aid-4f354c9c-2	artifact identity
Base commit	07cfcf662d978796aa4b7b46e8bccec27109effe	artifact currentness
Head commit	4f354c9cffed6d1e786c46102c974abf0fa4e1dd	artifact identity
Human-review lines	49	deterministic file stats
Canonical files	5	deterministic membership
Total files	13	deterministic membership
Tasks	1/1 done	source:t1
Acceptance	7/7 R-IDs evidenced	source:s-spec, source:t1
Full gate	run_tests_parallel.py green at the PR head (4991 ran, 0 failures); ruff 0.16.0 clean; sync-codex.sh idempotent (empty diff on the second run)	source:t1, source:d
Review	impl-review via codex: round 1 fan-out NEEDS_WORK (isolate concurrent children; commit the dirty remainder before inspection), round 2 SHIP	source:t1
Net	13 files, +384/-3; 5 canonical, 3 generated mirror, 5 Flow state	source:d

Change	Attention	File	Purpose	+/-	Diff	Evidence
MODIFIED	CANONICAL	plugins/flow-next/agents/worker.md	New Bridged implementer section (gate, who does what, keep/skip, pointer prompt, run the bridge, on return); Phase 1.5 skip note; Investigate-first rule names the child.	+19/-1	—	source:d
MODIFIED	CANONICAL	plugins/flow-next/skills/flow-next-work/phases.md	One paragraph after the implementer-tier sentence: worker bridges, conductor never does, TIMEBOX is the worker's.	+2/-0	—	source:d
MODIFIED	CANONICAL	scripts/sync-codex.sh	SECTION3C heredoc gains the implementer-tier and bridge paragraphs; two grep guards added to the dispatch-field check.	+7/-1	—	source:d
MODIFIED	CANONICAL	plugins/flow-next/skills/flow-next-work/references/no-plan-route.md	Bridged implementer paragraph appended to Judicious subagent use.	+18/-0	—	source:d
MODIFIED	CANONICAL	CHANGELOG.md	Unreleased Changed entry referencing #431 and #436.	+1/-0	—	source:d

Critical changes

• High-churn: .flow/specs/fn-245-worker-bridged-implementer-path-431.json (+149/-0 lines) — Flow state, not hand-written prose
• High-churn: .flow/specs/fn-245-worker-bridged-implementer-path-431.md (+82/-0 lines) — the spec this PR implements
• High-churn: .flow/memory/bug/integration/committing-bridged-children-cannot-2026-09-14.md (+25/-0 lines) — memory entry captured from the review fix round
• High-churn: .flow/tasks/fn-245-worker-bridged-implementer-path-431.1.md (+25/-0 lines) — task done summary and evidence
• High-churn: plugins/flow-next/agents/worker.md (+19/-1 lines) — the new worker contract

How to review this PR

The pipeline already verified this — you don't re-check it from scratch:

• Tests / gates: python3 scripts/run_tests_parallel.py green at the PR head (4991 ran, 0 failures); uvx ruff@0.16.0 check . clean; scripts/sync-codex.sh idempotent (empty diff on the second run)
• R-ID coverage: 7/7 acceptance criteria evidenced
• Cross-model review: impl-review via codex, round 1 fan-out NEEDS_WORK (isolate concurrent children; commit the dirty remainder before inspection), round 2 SHIP

Your job — the calls the pipeline can't make:

• Line-review the Must review bucket below — the one canonical file that carries the new contract.
• Spot-check the verified claims above; sample, don't re-run everything.
• Own what machines can't judge: whether the worker/child split of judgment and investigation is the one you want, and whether the prose earns its always-loaded context cost (G1).

Review plan

Must review (~5%)

• 🔴 plugins/flow-next/agents/worker.md — the only canonical file among the high-churn set; it is the new always-loaded worker contract — Does the section keep prompt composition, the diff check, the gates, review dispatch, and flowctl done with the worker while the child investigates, and does it forbid copying TIMEBOX into the child's brief? — open the ## Bridged implementer (gated on the implementer tier) section (also touched by c6a4872f)

Spot-check

• 🟡 plugins/flow-next/skills/flow-next-work/phases.md — one paragraph in step 3c: worker bridges, conductor never does, TIMEBOX is the worker's cap. Does it agree with the worker section word for word on who bridges?
• 🟡 plugins/flow-next/skills/flow-next-work/references/no-plan-route.md — the narrowed license (first draft in 85204f01, isolation rule added after review). Is the isolated-workspace-per-child rule the shape you want for parallel bridge calls?
• 🟡 scripts/sync-codex.sh — the mirror's 3c heredoc now carries the implementer-tier paragraph and the bridge paragraph, with two grep guards. Is restoring the paragraph the mirror dropped since Orchestration by intent: named tiers, reach pages, zero shipped slugs #352 wanted here, or a separate change?
• 🟡 CHANGELOG.md — one Unreleased entry; check the wording against the release-notes gate.

Safe to skim (~85%)

• ⚪ 3 files under plugins/flow-next/codex/ (agents/worker.toml, skills/flow-next-work/phases.md, skills/flow-next-work/references/no-plan-route.md) — regenerated by scripts/sync-codex.sh, guard-verified, idempotent on the second run — skim
• ⚪ 5 files under .flow/ (the spec .md/.json, the task .md/.json, the bug memory entry) — task-state, not hand-written code — skim

Memory left behind

Bugs captured during this spec:

• bug/integration/committing-bridged-children-cannot-2026-09-14 — The bridged-implementer prose in no-plan-route.md licensed "parallel bridge calls on disjoint surfaces, each with the same long-task brief" and then claimed "commit ownership and the join barrier stand unchanged".

---

Generated by /flow-next:make-pr from fn-245-worker-bridged-implementer-path-431 against main on 2026-09-14.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Gretn9Fz79EuiZXQkwnr1s

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-14T08:57:35.930246Z	4f354c9	PR opened

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" or "@codex security review".

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]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 4f354c9cff

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".

[chatgpt-codex-connector]

Link the worker to the packaged reach page

When this agent runs from an installed plugin in a consumer repository—the normal deployment—plugins/flow-next/docs/reach/ does not exist in that checkout: canonical agent links must resolve through ../docs/reach/README.md, while the Codex mirror installs the page beneath ../docs/flow-next/reach/README.md. Because the fresh-context worker receives only configuration fields and this new gate depends on the reach page to distinguish an in-host model from a CLI-only or unreachable one, it can silently select the wrong implementation path. Make this a relative Markdown link so sync-codex.sh rewrites and validates the installed layout.

AGENTS.md reference: AGENTS.md:L7-L8

Useful? React with 👍 / 👎.

[gmickel]

Closing in favour of #438, which shipped in 5.4.0 (https://github.com/gmickel/flow-next/releases/tag/flow-next-v5.4.0).

Your diagnosis was right and is adopted: step 3c routed implementation to the implementer tier but the worker never consulted it, so a pinned bridge model was silently ignored. #438 takes that fix (worker Phase 1b: pointer prompt plus the usage guide's brief, one foreground bridge call, dirty-tree commit, base..HEAD review, gates, then the normal review and done) and the 3c paragraph, with the mirror's implementer-tier paragraph restored.

Where it differs from this PR: fan-out stays with the bridged child rather than moving to the worker as parallel bridge calls in per-child worktrees. The direct route's premise is one owner with the whole task and its own delegation, so on the bridged path the owner is the child; the brief's never-list is back to five clauses (push, history rewrite, scope, verdict, nested bridge), and the July keep-the-child-flat caveat is now a dated watch after openai/codex#33267's own repro ran clean on codex-cli 0.153.4. The reasoning is recorded in STRATEGY.md (The owner holds the license) and criteria.md G3.

Thanks for the work here; the changelog credits the diagnosis.