autoconverge
Drives one draft PR to convergence in one autonomous run with a host-selected pacer. On workflow: three parallel lenses (code-review, bug-audit, self-review) and one-commit fixes via converge.mjs. On portable: pr-converge continuous ticks via portable-driver.md. Bugbot, Copilot, and Codex as terminal confirmation gates before ready. Use when the user says '/autoconverge', 'autoconverge this PR', 'converge this PR in one run', 'run the converge workflow', or 'drive the PR to ready autonomously'.
- License 许可
- MIT
- 最近更新
- 2026-11-15
- Stars
- ★ 7
- Forks
- 1
01 前言 Front Matter
- name
- autoconverge
- description
- Drives one draft PR to convergence in one autonomous run with a host-selected pacer. On workflow: three parallel lenses (code-review, bug-audit, self-review) and one-commit fixes via converge.mjs. On portable: pr-converge continuous ticks via portable-driver.md. Bugbot, Copilot, and Codex as terminal confirmation gates before ready. Use when the user says '/autoconverge', 'autoconverge this PR', 'converge this PR in one run', 'run the converge workflow', or 'drive the PR to ready autonomously'.
02 资源文件 50
- 参考文档.claude/CLAUDE.md14 B
- 参考文档AGENTS.md1 B
- 参考文档reference/.claude/CLAUDE.md14 B
- 参考文档reference/AGENTS.md1 B
- 参考文档reference/closing-report.md8.2 KB
- 参考文档reference/convergence-gates.md848 B
- 参考文档reference/convergence.md12.2 KB
- 参考文档reference/copilot-findings.md3.0 KB
- 参考文档reference/gotchas.md3.9 KB
- 参考文档reference/headless-safety.md2.4 KB
- 参考文档reference/multi-pr.md6.4 KB
- 参考文档reference/self-closing-loop.md4.3 KB
- 参考文档reference/stop-conditions.md8.8 KB
- 脚本test_portable_pacer_gate.py2.1 KB
- 参考文档workflow/.claude/CLAUDE.md14 B
- 参考文档workflow/AGENTS.md1 B
- 脚本workflow/aggregate_runs.py12.4 KB
- 脚本workflow/autoconverge_report_constants/__init__.py—
- 参考文档workflow/autoconverge_report_constants/.claude/CLAUDE.md14 B
- 参考文档workflow/autoconverge_report_constants/AGENTS.md1 B
- 脚本workflow/autoconverge_report_constants/render_report_constants.py17.5 KB
- 脚本workflow/converge_multi.mjs6.1 KB
- 脚本workflow/converge_multi.run-input.test.mjs6.4 KB
- 脚本workflow/converge.clean-audit.test.mjs31.8 KB
- 脚本workflow/converge.codex-gate.test.mjs21.1 KB
- 脚本workflow/converge.contract.test.mjs3.2 KB
- 脚本workflow/converge.copilot-gate.test.mjs47.0 KB
- 脚本workflow/converge.fix-progress.test.mjs4.5 KB
- 脚本workflow/converge.fix-recovery.test.mjs1.7 KB
- 脚本workflow/converge.merge-conflict.test.mjs797 B
- 脚本workflow/converge.mjs201.0 KB
- 脚本workflow/converge.p2-advance.test.mjs9.7 KB
- 脚本workflow/converge.path-aware.test.mjs2.1 KB
- 脚本workflow/converge.precatch.test.mjs10.6 KB
- 脚本workflow/converge.run-input.test.mjs4.8 KB
- 脚本workflow/convergence_summary.py3.8 KB
- 其他workflow/fixtures/wf_run/subagents/workflows/wf_881252e6-700/agent-a11d903476b803493.jsonl1.4 KB
- 其他workflow/fixtures/wf_run/subagents/workflows/wf_881252e6-700/agent-a26213978adeef6fb.jsonl865 B
- 其他workflow/fixtures/wf_run/subagents/workflows/wf_881252e6-700/agent-a3def0d15ed9d9110.jsonl4.1 KB
- 其他workflow/fixtures/wf_run/subagents/workflows/wf_881252e6-700/agent-a41f41b1b708ee3b7.jsonl414 B
- 其他workflow/fixtures/wf_run/subagents/workflows/wf_881252e6-700/agent-a758b880abecc3ff7.jsonl766 B
- 其他workflow/fixtures/wf_run/subagents/workflows/wf_881252e6-700/agent-a8897b89656b1bd16.jsonl872 B
- 其他workflow/fixtures/wf_run/subagents/workflows/wf_881252e6-700/agent-ab1c2d3e4f5a6b7c8.jsonl2.0 KB
- 其他workflow/fixtures/wf_run/subagents/workflows/wf_881252e6-700/agent-abd463d744a1437bc.jsonl412 B
- 其他workflow/fixtures/wf_run/subagents/workflows/wf_881252e6-700/agent-ad19d027ae8ee1816.jsonl1.4 KB
- 参考文档workflow/fixtures/wf_run/workflows/wf_881252e6-700.json6.4 KB
- 脚本workflow/render_report.py34.4 KB
- 脚本workflow/test_aggregate_runs.py4.6 KB
- 脚本workflow/test_convergence_summary.py4.6 KB
- 脚本workflow/test_render_report.py30.1 KB
03 SKILL.md 正文
Autoconverge
One launch drives the whole loop to convergence. The /autoconverge skill
scans the tool list, selects a pacer (Workflow or portable), then resolves
PR scope, enters a worktree, grants project permissions, and drives the shared
converge product to ready (or a named blocker).
| Pacer | Host surface | How the loop runs |
|---|---|---|
workflow |
Tool list includes Workflow |
converge.mjs background pass; state in the workflow journal |
portable |
Workflow absent |
Continuous in-session ticks per ../_shared/pr-loop/portable-driver.md; same helpers and check_convergence.py ready definition |
pr-converge selects schedule_wakeup or portable the same way. Both
entry skills share the helper scripts and the convergence gate.
Run scope: one PR or several
Decide the scope from how many PRs the user named, then follow that path:
- One PR → the single-PR run described below: one worktree, one pacer
launch (
converge.mjsonpacer=workflow, portable continuous ticks onpacer=portable), one teardown. - Several PRs → the Multiple PRs run: on
pacer=workflow,workflow/converge_multi.mjsdrives every PR in parallel; onpacer=portable, run the portable driver once per PR (serial or host fan-out), then one teardown per PR.
The single-PR sections (Requirements, Pre-flight, Run the workflow or portable driver, Teardown) each describe one converge run. The multi-PR reference reuses them once per PR and adds only what fanning out needs: a per-PR worktree and a per-PR teardown loop.
Requirements
Scan the tool list for Workflow and ScheduleWakeup, then select the pacer:
python "$HOME/.claude/skills/_shared/pr-loop/scripts/select_converge_pacer.py" \
--skill autoconverge \
--has-workflow <0|1> \
--has-schedule-wakeup <0|1>
pacer=workflow— continue with Pre-flight and Run the workflow below.pacer=portable— continue with Pre-flight (portable worktree rules whenEnterWorktreeis absent), then the continuous driver in../_shared/pr-loop/portable-driver.md. Do not abort because the Workflow tool is missing.
Set the GitHub transport directive before any GitHub step. Use transport: "cloud" in a
Cloud session. Use transport: "local" in a local session. Cloud uses the
connected GitHub MCP tools and never starts a local GitHub CLI authentication
flow.
Review-lens boundary
The code-review lens boundary (workflow agent versus the built-in /code-review
command) is defined on the Code-review lens bullet in
reference/convergence.md.
Select the GitHub transport directive before any GitHub step
In a Cloud session, set transport: "cloud" and load the GitHub MCP schemas
before the first GitHub tool call. Do not run any local GitHub CLI
authentication, account-switch, or token command. Use the operation in
cloud-pr-loop-compatibility.md
for each GitHub read or write. Use the Cloud session's credentialed push path
for local commits. If that path rejects a push, report the exact error instead
of trying to log in through the local CLI.
In a local session, set transport: "local" and use the authenticated local
GitHub CLI. The workflow defaults to this value when older callers omit the
field.
Pre-flight (main session)
-
Build the task list (step 1 of every run). After Copilot quota / down flags and Codex required/down are known, run only:
python "$HOME/.claude/skills/_shared/pr-loop/scripts/build_converge_task_list.py" \ --bugbot-down <0|1> --copilot-down <0|1> \ --codex-down <0|1> --codex-required <0|1>Register every
tasks[]entry on the session task list. Final task is alwaysall_runnable_reviews_clean_same_head(done_whenin the JSON). The run is complete only when that final task is completed — all runnable code reviews CLEAN on the same HEAD. Do not invent tasks in prose. Onpacer=portable,open-runreturns the same list; still register it as step 1 before driving reviews. -
Enter a worktree. When
EnterWorktreeis in the tool list, call it with no arguments before anygh,git, file read, or edit.gh/gitBash calls do not auto-isolate, so isolation is mandatory. If it fails, report and stop. A bareEnterWorktreebranches fromorigin/main; step 2 positions the worktree on the PR's head ref. WhenEnterWorktreeis absent, isolate with git worktree machinery per../_shared/pr-loop/portable-driver.md§ Isolation and worktree, then continue step 2. -
Resolve PR scope. When the user passed a PR URL or number, parse owner, repo, and number from it. Otherwise read the current branch's PR:
gh pr view --json number,headRefName,url,isDraft,baseRefName. Captureowner,repo,prNumber. Confirm the PR is a draft; if it is already ready, mark it draft first (gh pr ready <n> --repo <o>/<r> --undo) so the loop owns the ready transition.Position the worktree on the PR branch. The run reviews
git diff origin/main...HEADagainst this worktree's localHEADand pushes each fix to the PR branch, so the worktree sits on the PR's head ref at the PR HEAD before the loop starts. A worktree fresh offorigin/mainhasHEAD == origin/main, shows an empty diff, and reports a false convergence with zero findings. When a local worktree already tracks the PR branch, enter that one (pass its path toEnterWorktreewhen that tool exists; otherwisecdinto it). Otherwise put the checkout on the branch withgh pr checkout <number> --repo <owner>/<repo>(orgit fetch origin <headRefName>thengit switch <headRefName>). Confirm before launching:git rev-parse --abbrev-ref HEADequals the PR's head ref and localHEADequals the PR head SHA. -
Verify the worktree is the PR's repo (strict pre-flight). Run
python "$HOME/.claude/skills/_shared/pr-loop/scripts/preflight_worktree.py" --owner <owner> --repo <repo> --mode strict. It confirms the working directory is a checkout of the PR's own repo and thatgit worktreemachinery is healthy. A non-zero exit prints aPREFLIGHT_OUTCOMEline and anABORTline: report that line and stop. Autoconverge runs inside the PR's own repo, so a working directory rooted in a different repo (for example,claude-dev-envwhile the PR lives inllm-settings) or in no git checkout at all cannot continue. -
Grant project permissions. Apply the
pr-loop-lifecycleskill's Open section (../pr-loop-lifecycle/SKILL.md) — the grant command (grant_project_claude_permissions.py) and the auto-modeAskUserQuestionescalation for a blocked grant both live there. -
Copilot quota pre-check. Before the pacer starts the loop, apply the
reviewer-gatesskill's Copilot quota gate (../reviewer-gates/SKILL.md) once. Exit 0 maps tocopilotDisabled: false/copilot_down=false; any non-zero exit maps tocopilotDisabled: true/copilot_down=true, and the run skips the Copilot gate with no agent spawned. -
Branch on pacer. When
pacer=portable, follow../_shared/pr-loop/portable-driver.md§ Continuous tick loop and § Autoconverge entry on portable pacer, then skip Run the workflow (no Workflow tool). Whenpacer=workflow, continue with Run the workflow.
Run the workflow
(pacer=workflow only.) Call the Workflow tool against the colocated script:
Workflow({
scriptPath: "<this skill dir>/workflow/converge.mjs",
args: { owner: "<O>", repo: "<R>", prNumber: <N>, bugbotDisabled: false, copilotDisabled: false, homeDirectory: "<HOME>", transport: "<cloud|local>" }
})
scriptPath is the absolute path to workflow/converge.mjs inside this skill's
own directory (on this install,
<home>/.claude/skills/autoconverge/workflow/converge.mjs). homeDirectory is
the absolute path to the directory that holds .claude (resolve it once from
$HOME, or $env:USERPROFILE on Windows, with forward slashes). The workflow
runs in a sandbox with no access to environment variables, so the run reads the
home directory from this arg to build the path to the codex-review scripts the
Codex gate calls; leave it out and the Codex gate cannot find those scripts. Set
transport: "cloud" for Cloud and transport: "local" for local execution.
Set bugbotDisabled: true only when the user has opted Cursor Bugbot out for the
run; otherwise the workflow detects an opt-out or an unreachable Bugbot on its
own. Set copilotDisabled: true when the step 5 quota pre-check exits non-zero,
and false when it exits 0; on true the workflow skips the Copilot gate with
no agent spawned. The workflow runs in the background and notifies this session
on completion. Watch live progress with /workflows.
The moment the Workflow call returns its run id, write the durable handoff so a
fresh session can resume the same run:
python "$HOME/.claude/skills/_shared/pr-loop/scripts/write_handoff.py" \
--pr-number <N> --head-ref <branch> --phase workflow \
--resume-command "Workflow({scriptPath, resumeFromRunId: '<runId>'})" \
--run-id <runId>
Write it again when the result lands, so the handoff carries the final run id and names the teardown phase the fresh session picks up from.
The workflow returns
{ converged, rounds, finalSha, blocker, standardsNote, copilotNote, codexNote, cleanAuditNote, reuseNote, deferredPrs },
plus a userReview field on a blocker: "user-review" return. cleanAuditNote is
non-null when the environment refused the CLEAN bugteam review post and the run
recorded the bypass — see
reference/stop-conditions.md § Clean-audit post
bypassed.
deferredPrs is the list of draft environment-hardening PRs the standards-deferral
path opened this run, each as { owner, repo, prNumber, copilotDisabled, bugbotDisabled }.
The two flags carry this run's Copilot and Bugbot availability, so the next generation
skips a reviewer that is down or out of quota without re-probing. This list is the seed the
self-closing loop converges next.
Copilot findings — tier, verify, then route
The Copilot gate tiers each finding: a self-healing finding (style, type hints, imports, formatting, magic-value extraction, test-only or doc-vs-code fixes — nothing that changes observable runtime behavior) flows into the fix round with no user notification. A code-concern finding (logic, security, data handling, error-handling semantics, concurrency — the tier whenever in doubt) goes to a verification stage before any routing.
Each code-concern finding gets its own verifier agent, all in parallel, inside
the workflow. A verdict is conclusive only when an actual check ran: the verifier
executes a command against the flagged HEAD — running the code path with crafted
inputs, forcing the claimed error condition, or running a purpose-built test —
and captures its output. The verdict carries
{ verdict, checkCommand, checkOutput, evidence }; a conclusive verdict with an
empty checkCommand or checkOutput downgrades to inconclusive.
- confirmed — the check reproduces the defect. The finding becomes self-healing: it joins the fix round carrying its repro, and the fix re-runs that same check, adds a regression test where the suite covers the surface, lands in one commit, pushes, and replies on the thread with the fix SHA and the before/after output. No page.
- refuted — the check shows the code already behaves correctly in the exact scenario the finding claims is broken. The workflow replies on the thread with the command and output, resolves it, and counts it clean. No page.
- inconclusive — everything else, and the verifier's default: no runnable check exists, the check is infeasible here, the results are ambiguous, or the fix needs a product decision. Any doubt sorts here. Only inconclusive findings page the user.
A round whose code concerns all confirm or refute never returns
blocker: "user-review". On one or more inconclusive findings, the workflow
stops with converged: false, blocker: "user-review", and a userReview
field carrying
{ reviewUrl, findings: [{ file, line, severity, tier, title, evidence }] } —
evidence is the verifier's one-line note stating what check was attempted and
why it was not decisive.
The wait for a human belongs to the orchestrating session. On a
blocker: "user-review" return, run the
copilot-finding-triage skill for the ntfy
page only (the per-finding summary and evidence note plus the reviewUrl
Copilot review link), then hold for the user's response by pacer. On
pacer=portable, override triage's hold step — ntfy plus the portable hold
below; do not follow triage's ScheduleWakeup / send_later path.
pacer=workflow— triage hold: 45-minuteScheduleWakeup(orsend_laterwhen that is the host arm).pacer=portable— portable hold only: in-session poll or handoff per../_shared/pr-loop/portable-driver.md; neverScheduleWakeuporsend_later.
When the user answers within the window, follow their direction. When the window closes with no response, run normal teardown and report the inconclusive findings un-reviewed.
Budget stop
Branch the stop and resume path on the selected pacer. Full rule:
reference/stop-conditions.md § Budget stop.
pacer=workflow—converge.mjspaces against the workflowbudgetAPI and stops at a round boundary when a full round does not fit. On ablocker: "budget"return, write the durable handoff with the run id and theWorkflow({scriptPath, resumeFromRunId})resume command before stopping, so a fresh session resumes the paced run without the stopped session's transcript.pacer=portable— stop at a tick boundary when the session cannot cover a full clean tick; write the durable handoff and resume with/autoconverge <PR URL>(see../_shared/pr-loop/portable-driver.md). Do not instruct a Workflow resume on the portable path.
Teardown
pacer=portable
Skip Workflow-only report steps that need a workflow run id (journal merge,
HTML closing report, Artifact publish). Run pr-loop-lifecycle Close
(description rewrite when converged, working-tree clean, permission revoke)
and print the final report block using tick/check_convergence outcomes.
User-review holds use the portable in-session poll (or handoff) from
../_shared/pr-loop/portable-driver.md
— not ScheduleWakeup. Resume command: /autoconverge <PR URL>.
pacer=workflow (on workflow completion)
Teardown runs as an ordered checkpoint list. After each checkpoint finishes,
re-write the durable handoff with --phase teardown, the run id, and the
checkpoints done so far, so a fresh session that resumes reads handoff.json
completed_steps and skips the checkpoints already done:
python "$HOME/.claude/skills/_shared/pr-loop/scripts/write_handoff.py" \
--pr-number <N> --head-ref <branch> --phase teardown \
--resume-command "/autoconverge <PR URL>" \
--run-id <runId> \
--completed-steps "<checkpoints done so far>"
The checkpoints run in this order: report, close. Write the
handoff after each one finishes.
On teardown entry, when ~/.claude/runtime/pr-loop/bugteam-pr-<N>/handoff.json
exists, read its completed_steps and skip any checkpoint the list already
names, so a resumed run performs only the checkpoints left.
On a blocker: "user-review" return under pacer=workflow, the workflow held
one or more code-concern findings that stayed inconclusive after the
executed-check verification stage. Run the
copilot-finding-triage gate before
teardown: send the ntfy alert with the per-finding summary and evidence note and
the userReview.reviewUrl link, then hold with a 45-minute ScheduleWakeup (or
send_later when that is the host arm). Act on the user's direction when it
arrives inside the window; when the window closes with no response, fall through
to normal teardown and report the userReview.findings un-reviewed. The PR
stays a draft in this path — the workflow marked nothing ready. Under
pacer=portable, use the portable teardown hold above (ntfy plus in-session
poll or handoff; never ScheduleWakeup).
Before the checkpoints, when the workflow returned a non-null copilotNote
(the Copilot gate was bypassed), query the PR's reviews once more for a
copilot-pull-request-reviewer[bot] review on the final HEAD
(fetch_copilot_reviews.py, or the GitHub MCP pull_request_read method
get_reviews). Copilot typically posts within 10–15 minutes of a request, so
a review can land between the bypass and teardown. When one exists and
carries findings, mark the PR draft, route the findings through one fix
round per the pr-fix-protocol skill, re-verify, push, and mark the PR
ready again — then run the checkpoints.
-
When
convergedis true — build and publish the closing report. Skip this entire step (report, artifact publish, comment, Chrome open) when the workflow returned a non-nullblocker. Followreference/closing-report.md§ Building the report and § Publishing: resolve the seed journal from the run id, merge the PR's runs withaggregate_runs.py, spawn theconvergence-summaryagent (ageneral-purposesubagent) on the prompt it builds, draw the report withrender_report.py, publish the HTML via theArtifacttool, post the one marker-keyed<!-- autoconverge-report -->PR comment, and open the artifact URL in Chrome. After the report is published, write the handoff with--completed-steps "report". -
Close the run. Apply the
pr-loop-lifecycleskill's Close section (../pr-loop-lifecycle/SKILL.md): whenconvergedis true, rewrite the PR description and clean the working tree — seepr-loop-lifecycle/reference/teardown-publish-permissions.md§ Clean working tree and § Publish the final PR description; the workflow already marked the PR ready. The permission revoke always runs, including on a blocker exit. Write the handoff with--completed-steps "report,close". -
Print the final report:
/autoconverge exit: <converged | blocked> Rounds: <N> Final commit: <finalSha> Blocker: <blocker> # only when blocked Standards: <standardsNote> # only when a round deferred code-standard findings Copilot: <copilotNote> # only when Copilot was down or out of quota Codex: <codexNote> # only when the Codex gate was bypassed (codex_down or opt-out) Clean-audit: <cleanAuditNote> # only when the CLEAN bugteam post was bypassed Reuse: <reuseNote> # only when the reuse pass identified an improvement
What the workflow does each round
reference/convergence.md defines the whole loop:
the merge-conflict pre-flight, the reuse pass and its three landing criteria,
the round shape (static sweep, the three parallel reading lenses, dedup, the
one-commit fix flow, the standards-deferral path), the terminal Bugbot, Copilot,
and Codex gates, the per-role model tiers, the full-diff rule, and the exact
ready definition check_convergence.py enforces.
reference/stop-conditions.md lists every way
the run ends short of ready. Hard-won failure lessons live in
reference/gotchas.md.
Every agent prompt the workflow authors carries a headless-safety preamble —
read-only agents get a trimmed form, edit agents the full form with the rm
auto-allow paths — specified in
reference/headless-safety.md.
Multiple PRs
When the user names several PRs, run the multi-PR path in
reference/multi-pr.md: one worktree per PR
(git worktree add on each head ref, strict pre-flight per worktree, one
permission grant per repository, one Copilot quota check for the whole run),
then launch by pacer — pacer=workflow: one workflow/converge_multi.mjs
call with one entry per PR; pacer=portable: portable driver once per PR
(serial or host fan-out) — then single-PR teardown once per PR and a
one-line-per-PR summary.
Self-closing loop: converge the deferred PRs
Every autoconverge run — single-PR and multi-PR alike — ends with the
self-closing loop in
reference/self-closing-loop.md: the
orchestrating session converges the draft environment-hardening PRs the run
deferred (deferredPrs on a single-PR run, allDeferredPrs on a multi-PR
run), then the PRs those runs defer, generation by generation, until a
generation opens none. An empty seed list ends the loop at once. The reference
also carries the Conventional-Commit title rule each hardening PR must meet.
Folder map
SKILL.md— this hub.../_shared/pr-loop/portable-driver.md— portable pacer whenWorkflowis absent.workflow/converge.mjs— the convergence workflow script.workflow/converge_multi.mjs— the multi-PR fan-out driver: oneconverge.mjschild run per PR in parallel, each pinned to its PR worktree viarepoPath.workflow/aggregate_runs.py— merges every autoconverge journal for a PR into one journal and returns its deduped findings, fix summaries, round count, and final SHA.workflow/convergence_summary.py— builds the convergence-summary agent prompt over a PR's merged findings.workflow/render_report.py— builds the closing convergence insights HTML report, taking the summary from--summary-file.workflow/autoconverge_report_constants/— named constants for the report builder and the summary prompt.reference/convergence.md— the whole loop: reuse pass, round shape, terminal gates, model tiers, ready definition.reference/stop-conditions.md— every way the run ends short of ready, including the budget stop.reference/gotchas.md— hard-won failure lessons.reference/closing-report.md— the closing HTML report: data source, build steps, publishing.reference/multi-pr.md— the several-PRs path: per-PR worktrees, theconverge_multi.mjslaunch, per-PR teardown.reference/self-closing-loop.md— the deferred-PR generations and the Conventional-Commit title rule.reference/headless-safety.md— the agent-prompt preamble and the rm auto-allow paths.
04 同仓库更多 Skills
update
Fast-forwards a local repository's main branch to a remote's main after confirming the repo path and source remote through AskUserQuestion. Use when the user says '/update', 'update main', 'fast-forward main', 'sync main from origin', 'pull latest main into <path>', or 'bring main up to date'.
source-command-sr-loop
Converging cleanup loop: repeat /simplify until clean, then repeat /code-review --fix until clean
skill-builder
Skill lifecycle: classify, scaffold, write via the skill-writer-agent, self-audit, compose sub-skills, polish description triggers, enforce deterministic scripts. Triggers: build a skill, new skill workflow, improve this skill, optimize skill description, skill development lifecycle, skill modularity, skill composition, description trigger catalog, deterministic skill scripts.
session-tidy
Audit, clean, and consolidate session logs in the Obsidian vault. Fixes format drift, resolves orphaned next-steps, updates stale statuses, and generates project rollup summaries. Use when session logs feel cluttered or before starting a new project milestone. Triggers on '/session-tidy', 'tidy sessions', 'clean up session logs', 'session audit'.
session-log
Logs a session report as an HTML page, tracks vault context, extracts decisions, and outputs a /rename command. Use for /session-log, 'journal this session', 'log this work', 'session report', 'summarize/log/record this session', 'save session', 'capture session', or 'document what we did'.
run-claude-dev-env
Build, install, run, and test the claude-dev-env installer. Use when asked to run claude-dev-env, install the Claude Code config into ~/.claude, drive the installer, smoke-test it against a sandbox, or run its test suite.