Codex port of the Kimiflow feature/fix/release loop. Use automatically only for actionable implementation requests for substantial feature work that crosses product surfaces/subsystems, adds material integration/data/security/public-API/architecture risk, or needs real discovery. Discussion, ideation, recommendations, explanations, status requests, and wish formulations stay direct/read-only. Explicit Kimiflow always starts it; explicit direct or direkt always bypasses it. Do not auto-trigger...
Scanned 9/6/2026
Install to Claude Code
npx -y skills add kimikonapps/kimiflow --skill kimiflow --agent claude-codeInstalls into .claude/skills of the current project.
Are you the author of Kimiflow?
Add the live security badge to your README — it updates automatically with every re-scan.
[](https://www.skillsdirectory.com/skills/kimikonapps-kimiflow-6420f1d0)More formats (shields.io, HTML) on the badges page.
---
name: kimiflow
description: "Codex port of the Kimiflow feature/fix/release loop. Use automatically only for actionable implementation requests for substantial feature work that crosses product surfaces/subsystems, adds material integration/data/security/public-API/architecture risk, or needs real discovery. Discussion, ideation, recommendations, explanations, status requests, and wish formulations stay direct/read-only. Explicit Kimiflow always starts it; explicit direct or direkt always bypasses it. Do not auto-trigger for fixes, reviews, refactors, cleanup, docs/config, or small low-risk features."
---
# Kimiflow For Codex
Run the Kimiflow loop for the user's request.
Read installed `SKILL.md` once. Per phase call `active-run.sh phase-read --run .kimiflow/<slug> --phase <N> --file phases/<phase-file> --packet --write`; trust its packet and do not read phase/reference files separately. Never preload all `reference.md`. Then apply this host map.
## Routing
Apply the frontmatter routing boundary: implementation authority and a material trigger are required. Use the current target without a routing question; raw file count is insufficient. If no trigger is clear, work directly. Explicit overrides win.
## Invocation
Treat these as explicit Kimiflow requests:
- `$kimiflow`
- `@kimiflow`
- `$kimiflow --launcher` / `$kimiflow --menu`
- `$kimiflow full|grill|plan|build|quick|review|audit|fix|release`
- `$kimiflow <task>`
- `kimiflow full`, `kimiflow grill`, `kimiflow plan`, `kimiflow build`, `kimiflow quick`, `kimiflow review`, `kimiflow audit`, `kimiflow fix`, `kimiflow release`
- `run kimiflow ...`
- `with kimiflow ...`
## Host Map
Take the normalized absolute plugin-root path two levels above this installed `skills/kimiflow/SKILL.md`.
Copy that literal path into each helper command: `KIMIFLOW_HOST=codex /absolute/plugin/root/hooks/<helper>`.
Do not assign or expand `KIMIFLOW_PLUGIN_ROOT` inside the command; never use project-relative hooks.
Do not probe helpers with `--help` during a run; use the exact documented command shapes.
Apply the canonical Kimiflow workflow from `$KIMIFLOW_PLUGIN_ROOT/SKILL.md` with these Codex substitutions:
- `/kimiflow` in user-facing text means `$kimiflow` or an explicit "run Kimiflow" prompt in Codex.
- `/kimiflow`, `/kimiflow --launcher`, and `/kimiflow --menu` mean `$kimiflow`, `$kimiflow --launcher`, and `$kimiflow --menu` in Codex. Empty or vague explicit Kimiflow invocations open the context-aware launcher and must use `$KIMIFLOW_PLUGIN_ROOT/hooks/launcher-status.sh` for the status snapshot.
- `/kimiflow full|grill|plan|build|quick|review|audit|fix|release` maps directly to Codex. `full` forces the strict loop but no approval by itself; `grill|plan|review|audit` are no-code modes until later build authority/selection; `quick` proves product intent or asks one small batch and normally uses a no-worker Discovery pulse; `fix` diagnoses first and pauses only for a material decision; `review` maps to existing-feature/current-change review; `release` authorizes one project-profile release.
- `/kimiflow --project-map <quick|skip>` means `$kimiflow --project-map <quick|skip>` in Codex. Missing maps, per-section staleness checks, `coverage`-based Phase-2 depth (`compressed|targeted|full`), recommended-but-skippable delta refreshes, focus selection, storage targets, and Improve/Docs publishing use the same canonical Project Map rules and `hooks/project-map-status.sh`. Repo docs are publish-safe derivatives only; raw `.kimiflow/project/` maps and sensitive findings stay local/private unless the user explicitly overrides that policy.
- Kimiflow's Project-Map auto-refresh and lookup use the same `$KIMIFLOW_PLUGIN_ROOT/hooks/project-map-status.sh` in Codex with `KIMIFLOW_HOST=codex`. Phase 7 runs `project-map-status.sh refresh --changed` after verify (re-stamps touched sections to `current`, prunes deleted members, adopts new files by prefix with `NEW-FILE` hints, re-indexes `.sh` `symbols`); Map Bootstrap runs `index-symbols` after writing sections; Phase 2 runs `$KIMIFLOW_PLUGIN_ROOT/hooks/suggest-affected-sections.sh --intent <file>|--text "<terms>"` to rank candidate sections and feed their `paths` to `coverage --affected`. The non-blocking Stop nudge `$KIMIFLOW_PLUGIN_ROOT/hooks/map-staleness-nudge.sh` surfaces residual staleness (rate-limited, USER-visible `systemMessage`, never blocks). All are recommended-but-skippable; raw map facts stay local. → reference.md "Project Map Bootstrap".
- `/kimiflow --verify-feature <feature-or-path>` means `$kimiflow --verify-feature <feature-or-path>` in Codex. Existing feature checks run the canonical Phase-7 code-review ensemble from `reference.md` read-only over the named path or current diff: candidate lens output is verified by the Codex orchestrator before it is promoted to findings, and no code is edited.
- `kimiflow security scan` / `security diff|deep|ci-artifact|eval|promote` stays private/advisory and fail-closed.
- Large `full|plan` after final `confirmed`: one Sol `fork_turns="none"` worker replaces parent planning through Phase 4; the parent waits once for `READY_PHASE5|NEEDS_DECISION`. Phase 4 dispatches scheduled lenses in parallel on one frozen PLAN hash, seals and repairs the whole batch, and never resets the three-round cap.
- Phase-7 Review Ensemble runs every axis once per stable PLAN, then required repair deltas. Mandatory schema-4 saturation authenticates bounded bug-cascade evidence; only source-bound schema-3 repairs for that exact source are accepted. Protected impacts cannot be waived; stable classes and the absolute three-round-plus-zero-carry-closeout cap stop ping-pong.
- Active Session and Adaptive Execution contracts use `active-run.sh`; the originating task owns the run. Stop coalesces an explicit same-turn observation; `status`/`next-action` are read-only and `observe` records only decisive artifacts. For pending context rollover, use one fresh `top` worker with `fork_turns="none"`. Pass the literal workspace root, literal plugin root, exact allowed status/read/handoff commands, run/packet identity, retained artifact digests, phase, and named code; the worker must not rediscover helpers or inspect hook internals. Persist its result, then call `rollover-handoff` with the exact ID/digest. Shared-checkout edits require `conflict-check`=`allow_disjoint`; use canonical item/terminal helpers.
- `$KIMIFLOW_PLUGIN_ROOT/hooks/workspace-preflight.sh` routes: free Primary direct; busy/dirty gets ≤3 locked owned Fleet trees, then FIFO. Work in its returned root. Phase 3 `declare`s exclusive Primary/Fleet path-contract leases + `blocked_by` against PLAN; Phase 5 requires `write-gate` and post-advance `revalidate`. Commit, `integrate` the JSON-argv-checked combined candidate, then `needs-reconcile` or finish/`retire`. Never mutate manual/Codex trees or request routine commit/stash/clean; ask only about ambiguous foreign bytes and require `WORKING_TREE_GATE OPEN`.
- Kimiflow's clarify gate uses `$KIMIFLOW_PLUGIN_ROOT/hooks/clarify-gate.sh`. Fresh Contract-4 features preserve user language and require two explicit actions: discuss/accept scope, then correct/confirm the final flow. Generic assent never confirms; content-free receipts and the intent lock bind both stages. Contract-3/schema-1 resumes.
- Large/critical Codex runs must visibly spawn exactly one independent bounded Intent Critic with `fork_turns="none"` before writing `INTENT.md`; local judgment may never set `critic=passed`. “Do not parallelize implementation” does not disable this verification.
- The host hook records an exact intake action. On resume call `status` once and trust its stage/action; never inspect receipt-hook internals or manufacture a hook response. Continue Phase 1 after `scope_ready`; after `confirmed`, write intent and run the clarify gate.
- Before `confirmed`, do not call `append-item`, create/update the task widget, or patch phase progress; `phase-read` advances phase state mechanically.
- Kimiflow's Current-State Pulse / Gate uses `$KIMIFLOW_PLUGIN_ROOT/hooks/current-state-gate.sh` in Codex. Run it for every non-trivial run; only an explicitly changing/current external dependency, platform, API, or method is medium/high and requires a subject-bound fresh primary source. A local public API is low-risk by itself. Persisted schema-1 runs keep legacy recall.
- Discovery uses `$KIMIFLOW_PLUGIN_ROOT/hooks/discovery-gate.sh`; `$KIMIFLOW_PLUGIN_ROOT/hooks/codebase-basis.sh` first binds HEAD and affected-path bytes/types. Prove `reuse → evolve → new`, compare research to that basis, and reject scope expansion.
- Conformance uses `$KIMIFLOW_PLUGIN_ROOT/hooks/conformance-gate.sh`. Bind up to five material decisions to evidence/paths/ACs/checks and `review_only|spike_required|runtime_required`; required spikes bind executed fixture, command, and output. Trace Contract-4 requirements like Contract 3.
- Decision-bearing artifacts, findings, learnings, and final reports load all of `$KIMIFLOW_PLUGIN_ROOT/references/workflow-prose-quality.md` only at their existing writing/review boundary and apply it in the same model pass; it adds no agent, model call, step, or gate.
- Kimiflow's fix-mode Red-Green Gate uses `$KIMIFLOW_PLUGIN_ROOT/hooks/red-green-gate.sh` in Codex. A fix run records Red/Green/Regression evidence in `BUG-REPRO.md`; `RED_GREEN_GATE OPEN` is required before Phase 7, learning promotion, or `Status: done`.
- Kimiflow's lazy frontend lane uses `$KIMIFLOW_PLUGIN_ROOT/hooks/frontend-quality-gate.sh`. Phase 0 records only the clean Contract-1 start; Phase 2 loads `$KIMIFLOW_PLUGIN_ROOT/references/frontend-quality-standard.md` and the optional Flagship delta only for UI features/polish; Phase 6 loads the plugin-rooted QA file and requires `FRONTEND_QUALITY_GATE OPEN`; Phase 7 rechecks the serialization preflight. Fix/audit/off and legacy/read-only runs never receive design payloads or synthetic screenshots.
- Kimiflow's local diagnostics advisory uses `$KIMIFLOW_PLUGIN_ROOT/hooks/lsp-diagnostics.sh` in Codex. It runs a bounded set of existing local diagnostics tools or one untracked `.kimiflow/lsp-diagnostics` command, never installs anything, rejects free-form CLI commands, classifies `FLAG`s by changed-file relevance, and routes them to `ADVISORIES.md`.
- Kimiflow's Memory Router uses `$KIMIFLOW_PLUGIN_ROOT/hooks/memory-router.sh`. Keep canonical bounded local-first recall/history/outcome/provider/Vault behavior; a prior-work cue replaces broad recall/Vault Pulse with one targeted query. Confirmed root causes may cite up to eight validated `Source evidence:` paths whose byte drift stales the learning.
- `${CLAUDE_PLUGIN_ROOT:-$CLAUDE_SKILL_DIR}` means the installed Kimiflow plugin root. In Codex, use `KIMIFLOW_PLUGIN_ROOT`.
- When invoking Kimiflow helper scripts from Codex, set `KIMIFLOW_HOST=codex`.
- `TaskCreate` / `TaskUpdate` means use Codex's task plan/status updates.
- Claude Code subagent names map to Codex subagents as follows:
- bounded file/symbol/map/log lookup: use a Codex `explorer` with `model: gpt-5.6-luna` and `reasoning_effort: low|medium`; the `top` model still performs Phase-2 synthesis.
- normal implementation or fix worker: use a Codex `worker` with `model: gpt-5.6-terra` and `reasoning_effort: medium|high`; promote to Sol for a named high-risk trigger.
- planning, review, independent semantic verification, risky diagnosis, or general quality work: use a Codex `default` with `model: gpt-5.6-sol` and `reasoning_effort: high|xhigh` unless a `cross_family_top` seat applies.
- Codex capability mapping is `top=gpt-5.6-sol`, `balanced=gpt-5.6-terra`, `cheap=gpt-5.6-luna`. Prefer Sol for the active Kimiflow session. If Codex exposes a Terra/Luna session, record the quality fallback and continue without a model-switch prompt; use deliberate Sol review/verification seats when available. Never use `ultra` inside Kimiflow because it adds nested automatic delegation under Kimiflow's own orchestrator.
- Keep Frontier prompts lean: goal, user boundaries, relevant evidence/tools, output contract, and exit criteria. Avoid provider folklore, forced reasoning, or generic checklists. A model/profile/prompt change gets a new fingerprint and must recalibrate adaptive savings.
- Kimiflow's cross-family seats (→ canonical `reference.md` "Model routing (per-role)") use the `claude` CLI on the Codex host: attempt condition `command -v claude`, invocation `claude -p --model fable --effort high "<prompt>"` (the final message is stdout). This pins the current strongest Claude tier; unavailable/refused/limited calls use the normal sticky same-family fallback. Never inherit an unverified default/cheap tier.
- `WebSearch` / `WebFetch` means Codex web/search or another available current-source tool. For current external technical facts, prefer primary sources.
- `CLAUDE.md` is a Claude project convention file. In Codex, read `AGENTS.md` first, and also read `CLAUDE.md` if it exists because Kimiflow historically treats it as a conventions hint.
## Gate Commands
Use the bundled scripts as the only mechanical source of truth:
- `$KIMIFLOW_PLUGIN_ROOT/hooks/resolve-review-gate.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/plan-review-gate.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/plan-blocker-gate.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/conformance-gate.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/resolve-build-gate.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/resolve-verbosity.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/working-tree-gate.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/workspace-preflight.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/clarify-gate.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/codebase-basis.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/current-state-gate.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/review-convergence-gate.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/discovery-gate.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/red-green-gate.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/frontend-quality-gate.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/lsp-diagnostics.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/launcher-status.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/active-run.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/memory-router.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/test-weakening-scan.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/secret-content-scan.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/project-map-status.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/suggest-affected-sections.sh`
- `$KIMIFLOW_PLUGIN_ROOT/hooks/map-staleness-nudge.sh`
For Codex invocations, call them with `KIMIFLOW_HOST=codex`, for example:
```bash
KIMIFLOW_HOST=codex "$KIMIFLOW_PLUGIN_ROOT/hooks/resolve-review-gate.sh" .kimiflow/<slug>/findings --round 1 --expect code-verified
```
Do not replace these scripts with model judgment. If a resolver says the gate is closed, the gate is closed.
## Output
Reply in the user's language. Keep Kimiflow's terse output rule from the canonical workflow: visible chat is control-plane only; artifacts and evidence go to files.
Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.
No comments yet. Be the first to comment!