Hold loop reports to the app's criteria at post time and keep phase prompts out of New Session

AI Assisted
This commit is contained in:
jparkerweb committed 2026-11-03 04:08:36 -08:00
1 parent 33f1783979
commit 61f27f81ab
7 files changed
+497 -35

No files matched your search

+1 -1
View File
@@ -42,7 +42,7 @@ Publish it with `post --wait`. A question stops the build: never post `"status":
| A spec gap only they can close | `text` | |
| Something learned for AGENTS.md | `confirm` | proposed lines in `consequences` |
Titles stay plain: "Which phase next", "Tests are failing", "Keep going?". In loop mode (`notes/loop.md`) a question is held or stopped by the app, never answered by it; the findings pick and fixes are `handledBy:"loop"` cards.
Titles stay plain: "Which phase next", "Tests are failing", "Keep going?". In loop mode (`notes/loop.md`) a question is held or stopped by the app, never answered by it; the findings pick and fixes are `handledBy:"loop"` cards. A `loopOutcome` refused at post time names the missing piece against the app's `loopExpect`: do that step and post again.
## Stop during a build
+9
View File
@@ -177,6 +177,15 @@ At the end of the run, post once a top-level `loopOutcome`:
- `subject` is one imperative line under 100 characters saying what changed (it
becomes the commit subject). `summary` is one line under 200 characters.
The app stamps `loopExpect` into the session as the phase starts: the criteria
it will judge the report by (`review`, `fix` as `"smart"` or the level list,
`guess`). It is the app's field, never yours to post. The console holds every
`loopOutcome` to it, so a report missing a required piece (`review.ran: false`
on a review run, no `review.triage.ran` under Smart pick, a skipped Critical in
the triage, no re-check with fixes chosen, assumed tasks without `--guess`) is
refused at post time with the field named, while the run can still do that
step and repost.
Then post `{"agent":{"status":"waiting"}}` and keep waiting with the usual
bounded `wait` slices until the app ends the session. Never end the turn on text.
If the post is refused (exit 3), fix the field it names and post again.
+84 -10
View File
@@ -518,6 +518,9 @@ export function applyPatch(state, patch) {
if (Object.hasOwn(patch, "loop")) throw new Error(rule("", "loop", "is set by open --loop, not by a patch"));
for (const [k, v] of Object.entries(patch)) {
if (v === null) {
// The loop's report fields are set, never deleted: a null here would
// strip the app's criteria or an already-posted report.
if (["loopExpect", "loopOutcome", "loopEnd"].includes(k)) throw new Error(rule("", k, "cannot be deleted by a patch"));
delete next[k];
continue;
}
@@ -578,6 +581,20 @@ export function applyPatch(state, patch) {
if (state.loopOutcome !== undefined) throw new Error(rule("", "loopOutcome", "was already posted for this session"));
next.loopOutcome = v;
break;
case "loopExpect":
// The app's judging criteria for the run's report: a `loopOutcome` is
// held to them in validate(), so a report missing a required piece is
// refused here while the agent can still fix it. The app owns the
// field and stamps once per session (an unlock or a hop clears it
// first); a second stamp, from anyone, is refused while the first is
// valid, but a corrupt stored stamp can be overwritten. An early
// self-stamp only weakens this early check: the driver's verdict is
// unchanged.
if (!state.loop?.driving) throw new Error(rule("", "loopExpect", "is only accepted in a loop session"));
if (isPlainObject(state.loopExpect) && loopExpectProblems(state.loopExpect).length === 0) throw new Error(rule("", "loopExpect", "was already stamped for this session"));
if (state.loopOutcome !== undefined) throw new Error(rule("", "loopExpect", "came after the run's report; post it while the phase builds"));
next.loopExpect = v;
break;
case "loopEnd":
// The app's word on how the loop ended; a later one replaces the earlier.
next.loopEnd = v;
@@ -826,8 +843,10 @@ export function validate(state) {
}
}
// The loop's two reports: the skill's outcome and the app's ending.
if (state.loopOutcome !== undefined) problems.push(...loopOutcomeProblems(state.loopOutcome));
// The loop's reports: the app's criteria, the skill's outcome and the
// app's ending.
if (state.loopExpect !== undefined) problems.push(...loopExpectProblems(state.loopExpect));
if (state.loopOutcome !== undefined) problems.push(...loopOutcomeProblems(state.loopOutcome, state.loopExpect));
if (state.loopEnd !== undefined) problems.push(...loopEndProblems(state.loopEnd, state));
// A note on a card is unanswered the moment no agent words follow it in its
@@ -1067,13 +1086,29 @@ const isPosInt = (n) => Number.isInteger(n) && n > 0;
// One line of 1..max characters.
const oneLine = (s, max) => typeof s === "string" && s.length >= 1 && s.length <= max && !/[\r\n]/.test(s);
/**
* Problems with a `loopExpect` (the app's judging criteria for the run's
* report), as sentences naming the field; an empty array when it is good.
*/
export function loopExpectProblems(e) {
const p = [];
if (!isPlainObject(e)) return ["loopExpect must be an object"];
if (typeof e.review !== "boolean") p.push("loopExpect.review must be true or false");
if (e.fix !== "smart" && !(Array.isArray(e.fix) && e.fix.every((l) => LOOP_LEVELS.includes(l)))) {
p.push(`loopExpect.fix must be "smart" or an array of ${LOOP_LEVELS.join(", ")}`);
}
if (typeof e.guess !== "boolean") p.push("loopExpect.guess must be true or false");
return p;
}
/**
* Problems with a `loopOutcome` (what the skill reports after building a phase
* under the loop), as sentences naming the field; an empty array when it is
* good. Rules are the overview's API section; house rules 4 and 5 apply to
* every string.
* every string. `expect` is the app's `loopExpect`, when the session holds
* one: the report is then also held to what this loop asked for.
*/
export function loopOutcomeProblems(o) {
export function loopOutcomeProblems(o, expect) {
const p = [];
const bad = (field, rule) => p.push(`loopOutcome.${field} ${rule}`);
if (!isPlainObject(o)) return ["loopOutcome must be an object"];
@@ -1178,6 +1213,43 @@ export function loopOutcomeProblems(o) {
if (!oneLine(o.subject, 100)) bad("subject", "must be one line of 1 to 100 characters");
else prose("subject", o.subject);
// A report that misses what the app said it would judge by is refused here,
// while the run can still do the step and repost, rather than stopping the
// phase on the driver's verdict later. The driver checks the criteria only
// on a "passed" report, so a "failed" one always lands (its reason is the
// phase's own word).
if (isPlainObject(expect) && o.result === "passed") {
if (isPlainObject(o.tests) && o.tests.ran === true && o.tests.passed === false) {
bad("tests.passed", "must be true on a passed report: the app refuses failed tests");
}
if (expect.review === true && isPlainObject(o.review)) {
if (o.review.ran === false) bad("review.ran", "must be true: this loop asked for a review");
if (expect.fix === "smart") {
if (!isPlainObject(o.review.triage) || o.review.triage.ran !== true) {
bad("review.triage.ran", "must be true: Smart pick requires the triage and its skipped list in the report");
} else if (Array.isArray(o.review.triage.skipped) && o.review.triage.skipped.some((u) => isPlainObject(u) && u.level === "critical")) {
bad("review.triage.skipped", "must not hold a skipped Critical: the app stops the phase on one");
}
}
const chosen = expect.fix === "smart" ? LOOP_LEVELS : Array.isArray(expect.fix) ? expect.fix : [];
if (chosen.length > 0 && o.review.recheck === false) {
bad("review.recheck", "must be true: findings at the loop's chosen levels need the re-check");
}
// The driver's `leftover()`: the counter and the unfixed list, whichever
// is larger; a leftover at a chosen Critical or Warning fails the phase.
const left = (level) =>
Math.max(o.review.left?.[level] || 0, (Array.isArray(o.review.unfixed) ? o.review.unfixed : []).filter((u) => isPlainObject(u) && u.level === level).length);
for (const level of ["critical", "warning"]) {
if (chosen.includes(level) && left(level) > 0) bad(`review.left.${level}`, "must be 0: a leftover at a chosen level fails the phase");
}
}
// The report's own claim only: the `[?]` marks in the spec stay the
// driver's call (`specs.mjs status`), which this post-time check cannot see.
if (expect.guess === false && isPlainObject(o.tasks) && Array.isArray(o.tasks.assumed) && o.tasks.assumed.length > 0) {
bad("tasks.assumed", "must be empty: this loop does not allow best guesses");
}
}
return p;
}
@@ -1225,9 +1297,10 @@ export function loopEndProblems(e, state) {
return p;
}
// Loop mode adds three optional keys that a blank state does not hold (absent,
// not null): `loop: { driving: true }` (set by `open --loop`), `loopOutcome`
// (the skill's report) and `loopEnd` (the app's).
// Loop mode adds four optional keys that a blank state does not hold (absent,
// not null): `loop: { driving: true }` (set by `open --loop`), `loopExpect`
// (the app's judging criteria), `loopOutcome` (the skill's report) and
// `loopEnd` (the app's ending).
export function blankState({ sid, workflow, title, project, specDir }) {
return {
schema: SCHEMA,
@@ -1261,9 +1334,10 @@ export function blankState({ sid, workflow, title, project, specDir }) {
*/
export function switchWorkflow(state, { workflow, title, specDir }) {
const next = { ...blankState({ sid: state.sid, project: state.project, workflow, title, specDir }), created: state.created };
// A blank state holds no `loop`, `loopOutcome` or `loopEnd`, so a hop never
// carries them over; `open --loop` stamps `loop` again after the switch.
for (const key of ["loop", "loopOutcome", "loopEnd"]) delete next[key];
// A blank state holds no `loop`, `loopExpect`, `loopOutcome` or `loopEnd`,
// so a hop never carries them over; `open --loop` stamps `loop` again after
// the switch, and the app restamps `loopExpect` on its next look.
for (const key of ["loop", "loopExpect", "loopOutcome", "loopEnd"]) delete next[key];
if (state.chat) next.chat = state.chat;
return next;
}