Skip to content

Mola-maker/mstack / skills/m-goal

m-goal

Verified

Long-running goal loop: goal file, checkpoints, stop rules, mandatory closeout. Use for run-until-done work, 别停/一直做到完成.

m-goal — 长跑目标循环

m-goal drives a long-running objective to completion across checkpoints and session restarts, converging on the target outcome rather than pausing for approval. It keeps a durable goal file so a resume never loses work, and it fills that file's ## Closeout section on every single exit path — done, failed, interrupted, or budget exhausted.

When to use / not use

Use for "keep going until done", multi-phase or multi-session work, or anything the user walks away from ("going to bed", "trust it when I'm back"). Do not use for a single focused fix, a bounded plan (use m-plan then m-dispatch), or pure planning with no execution loop. If the goal is an open-ended optimization target — a metric like performance, score, or pass rate that should keep climbing — hand the loop to m-team, which re-forms an agent team round by round; m-goal still owns the goal file and writes the closeout when m-team returns. If work is reversible, proceed and report rather than asking.

Procedure

Using m-goal: — say this first, before any tool call.

Rule: no done status without a written closeout, a passing gate, and at least the minimum rounds. No exceptions: do not write the closeout "after one more try", in a second file, or from memory of the run; do not loosen a criterion to reach it (amend it, step 6); do not stop on a first lucky pass; do not replace the counter script with your recollection.

You will be tempted to think…Reality
"I'll write the closeout when I've finished the last fix."The loop dies where you do not expect (a freeze hook denied the write; the session ended). The closeout is written on the exit path, not after it (reviews/2026-11-06-round2.md #10).
"The state script is not at that path, so I'll count failures myself."A memory count is no count; the 3-in-a-row rule then does not exist (round2 #1). Find the script from the skill dir, or do it by hand and say so in the report.
"The suite is red already, I'll save the baseline after my first edits."A late baseline bakes your own failures in; no baseline blocks the task forever (round2 #6). Save it before the first edit.
"The review had one unconfirmed finding, so the review failed." (or "so it is fine")Neither: pass means no confirmed in-scope finding and no axis issues (round2 #7). Anything else is a Ruling line.
"The metric hit target once, we are done."One lucky run ends the loop early; the stop predicate is target AND min-rounds (research/round3/pstack-craft.md T5).
"The closeout can go in its own file."Resume reads only the goal file and re-does a finished goal (reviews/2026-11-04-acceptance.md #4).

Red flags — you are breaking the skill now: you typed "done" with no ## Closeout status line; the state file's rounds is below the minimum and you are writing done; you are on a fourth attempt at the same signature; you rewrote a checkpoint row; you edited an acceptance line in place.

  1. Write the goal file. Create .mstack/goals/<slug>.md (layout: references/goal-file-format.md) containing: the objective in one testable sentence, the ordered sub-goals (outcome first), each sub-goal's acceptance criteria, the per-task attempt cap, the minimum rounds (default 3; a number you set before starting), the stop conditions, the Boundaries line (what the run may touch) and the Output & approver line (where the result lands, who approves irreversible actions), and empty ## Amendments / ## Closeout placeholders. A goal that repeats one action over many items runs 3 items first, reads the results, then runs the rest. Break the brief into the smallest sub-goals that each end in a check (principle 3). Then, before the first edit: save the m-verify baseline — run the suite once and write one signature line per already-failing test to .mstack/verify/baseline.txt (format in m-verify's references/verify-gates.md; a green suite means an empty file) — and record the commit that baseline belongs to: node "<this skill dir>/scripts/goal-state.mjs" init .mstack/goals <slug> --start-commit <sha> --min-rounds <N> <this skill dir> is the directory this SKILL.md was loaded from (Claude Code shows it as the skill base directory). If the script cannot be found, do this step by hand and say so in the report. <sha> is git rev-parse HEAD, or none when the repo has no commits.
  2. Order by risk. Architectural decisions and core abstractions first, integration points next, spikes/unknowns, then polish last — fail fast on risky work.
  3. Loop per sub-goal. Pick the first not-done sub-goal; implement; verify each acceptance criterion with fresh evidence (run it, read the output — principle 4); mark done only when all criteria pass. Spend extra budget on several attempts plus selection between them, never on one bigger single pass; at a task's attempt cap, route around it or escalate.
  4. Checkpoint after each sub-goal. Append one row to the checkpoint table — fixed 5 columns ts / decision / why / evidence / result, append-only (a wrong entry is superseded by a new row). Then update the loop state with this exact command: node "<this skill dir>/scripts/goal-state.mjs" record .mstack/goals <slug> [--round] [--fail <signature>] [--progress] [--tokens N] [--budget PCT] [--status active|done|failed] Pass --round once per completed verify pass over the acceptance criteria (a round is that, not a record call); pass --fail <signature> for every error signature that failed this round and --progress only when a sub-goal verifiably advanced (result: pass); --progress resets the consecutive counters. --tokens and --budget are your own estimate against the cap in the goal file. The script prints ok / warn / stop plus rounds=N min=M, and exits 1 on stop, including when --status done is rejected for rounds below the minimum; read it every time. Exit 2 means the state file is unreadable: repair it by hand, never delete it to reset the counters.
  5. Final gate, fixed order. Only when every sub-goal passes and rounds is at least the minimum: m-deslop on the changed files only, then m-verify, then m-review. A failed review appends a new sub-goal and the loop continues.
  6. Write the closeout (mandatory). Fill ## Closeout in the same goal file with the fields in references/closeout.md (including the gate block and Attention), then stop. A criterion proven wrong is amended there, never edited in place.

Ruling record: any decision this skill does not cover becomes one line Ruling: decision | reason | cost if wrong in the checkpoint table (see references/closeout.md). Reversible decisions never wait for the user: take the default, record it, continue.

Gate function — before writing done or any success claim:

claimevidence requiredevidence that is NOT enough
objective meteach acceptance command run this round, output readthe diff looks right; an earlier round passed
loop may stoptarget met AND rounds ≥ minimum in the script lineone lucky pass; your own tally
gate passedone proving command per m-deslop / m-verify / m-review"no complaints"; a sub-agent said ok

Output

.mstack/goals/<slug>.md (goal, sub-goals, checkpoints, ## Closeout) plus .mstack/goals/<slug>.state.json (loop counters). Closeout fields, in order, and the amend rule are in references/closeout.md. Report in the standard shape: one-line conclusion → what changed with evidence → what was not done / risks. If the outcome was done, mark all sub-goals complete and note the reviewer verdict; if not, leave the goal file intact so a later resume or m-handoff continues from the last checkpoint. Never claim completion without a written closeout. m-handoff only links to the closeout; it never writes one.

Stop conditions

Write the closeout and stop when: the objective is met AND rounds ≥ minimum; the user says stop/cancel; one error signature fails 3 times in a row; or the budget reaches 100% (warn at 90%). Fewer rounds than the minimum with the objective met means run another verification round, not stop. Ask the user only for an irreversible, external, or genuinely-undecidable choice (principle 8). After the third identical failure change the hypothesis (principle 9) — there is no fourth try.

Harness notes

  • Claude Code: a persistent goal can pair with its native goal mechanism, but state lives in .mstack/goals/ so it survives restarts; DSH and Codex both read/write the same file.
  • Budget cap: DSH max_goal_rounds or a token budget maps to the budget-exhausted closeout; none of the harnesses auto-write it, so m-goal must write it before exiting.
  • Headless (Codex/DSH) runs have no one to ask: record every guess as an assumption in the goal file rather than waiting.
  • scripts/goal-state.mjs needs Node (zero dependencies). Without Node, edit .mstack/goals/<slug>.state.json by hand — the schema is in references/goal-file-format.md — and still apply the 3-in-a-row / 90% / 100% rules. On closeout, set the file's status to done or failed (--status, or by hand) so a stale check can tell this goal is finished.
slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack."Stars: 198kUpdated Claude CodeCodexVerifiedInstall
ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt.Stars: 196kUpdated Claude CodeCodexVerifiedInstall