Rehome the goose-ask and aos-role-question probe notes #1126

New issue

Open

opened 2026-08-18 15:59:21 +00:00 by coilyco-ops · 2 comments

coilyco-ops commented

2026-08-18 15:59:21 +00:00

Member

PR #1124 evicted docs/test-harness.md to free a slot under the large-band 40-doc cap for docs/aos-eval.md. That page was a parent index for a test-harness-<agent> family that was never authored, with its one Goose child stapled inline.

Two live things it documented have no page now:

goose-ask (scripts/goose-ask.sh, ward verb) - the one-shot probe wrapper, its banner stripping and transcript path under ~/.cache/agentic-os/goose-ask/, the -f/-s/-m flags, and the note that ward gates repo verbs on a clean tree so the raw script is the iteration path.
aos-role-question - the composed-role probe, its cloud and local mechanisms, the tmpfs storage argument, the ten-minute ceiling, and the ROLE-CONFIRMED: <role> output requirement.

Both still work and both are described in the justfile. The content is at the parent of the #1124 commit if it should come back.

Decide one of: fold into docs/aosguard.md or docs/aos-cli.md (both are near the size cap, so something else moves), give the probes their own page and evict another (the docs count is at 40 again), or accept the justfile descriptions as sufficient and close this.

PR #1124 evicted `docs/test-harness.md` to free a slot under the large-band 40-doc cap for `docs/aos-eval.md`. That page was a parent index for a `test-harness-<agent>` family that was never authored, with its one Goose child stapled inline. Two live things it documented have no page now: * `goose-ask` (`scripts/goose-ask.sh`, ward verb) - the one-shot probe wrapper, its banner stripping and transcript path under `~/.cache/agentic-os/goose-ask/`, the `-f`/`-s`/`-m` flags, and the note that `ward` gates repo verbs on a clean tree so the raw script is the iteration path. * `aos-role-question` - the composed-role probe, its cloud and local mechanisms, the tmpfs storage argument, the ten-minute ceiling, and the `ROLE-CONFIRMED: <role>` output requirement. Both still work and both are described in the justfile. The content is at the parent of the #1124 commit if it should come back. Decide one of: fold into `docs/aosguard.md` or `docs/aos-cli.md` (both are near the size cap, so something else moves), give the probes their own page and evict another (the docs count is at 40 again), or accept the justfile descriptions as sufficient and close this.

coilyco-ops commented

2026-08-18 16:24:50 +00:00

Author

Member

What test-harness.md carried that is still live and now has no page: goose-ask's transcript path, batch and system-text flags, and the clean-tree gotcha that makes the raw script the iteration path, plus aos-role-question's cloud and local mechanisms, the tmpfs storage argument, the ten-minute ceiling, and the ROLE-CONFIRMED: <role> requirement.

What was genuinely legacy on it: a test-harness-<agent> parent index for a family never authored, with its one Goose child stapled inline.

The wall is the 40-doc cap for the large band, which the repo was already at before docs/aos-eval.md. docs/aos-eval.md has about 8 lines and 1900 chars of headroom, nowhere near the ~40 lines this needs, and folding probe docs into it would break its one-subject scope anyway.

So restoring the page means evicting a different doc, and my survey found no dead one: the thinnest candidates (aos-roles-and-voice.md, context-budget-roles.md) are all load-bearing or referenced from source. Options stay as listed above, with one added: restore test-harness.md minus its never-authored parent scaffold, and evict whichever doc Kai judges least load-bearing.

Reframing this after Kai's correction on agent-compose#305: the instruction was to supersede the legacy parts of a page, not to remove the page. So the question here is not only where the notes go, it is whether the eviction should stand at all. What `test-harness.md` carried that is still live and now has no page: `goose-ask`'s transcript path, batch and system-text flags, and the clean-tree gotcha that makes the raw script the iteration path, plus `aos-role-question`'s cloud and local mechanisms, the tmpfs storage argument, the ten-minute ceiling, and the `ROLE-CONFIRMED: <role>` requirement. What was genuinely legacy on it: a `test-harness-<agent>` parent index for a family never authored, with its one Goose child stapled inline. The wall is the 40-doc cap for the large band, which the repo was already at before `docs/aos-eval.md`. `docs/aos-eval.md` has about 8 lines and 1900 chars of headroom, nowhere near the ~40 lines this needs, and folding probe docs into it would break its one-subject scope anyway. So restoring the page means evicting a different doc, and my survey found no dead one: the thinnest candidates (`aos-roles-and-voice.md`, `context-budget-roles.md`) are all load-bearing or referenced from source. Options stay as listed above, with one added: restore `test-harness.md` minus its never-authored parent scaffold, and evict whichever doc Kai judges least load-bearing.

coilyco-ops commented

2026-08-18 17:00:07 +00:00

Author

Member

Resolved on the branch for #1124, without an eviction. The content is back in docs/aos-eval.md as a closing "probe layer underneath" section: goose-ask's transcript path, its -f/-s/-m flags, the clean-tree gotcha that makes the raw script the iteration path, and aos-role-question's cloud and local mechanisms, the tmpfs argument, the ten-minute ceiling, and the ROLE-CONFIRMED requirement.

Folded rather than given its own page for the reason above: docs/*.md is at the 40-doc cap, every other page is within a few lines of its own, and no pair merges under 120. The one candidate that looked splittable, aos-roles-and-voice.md, is 53 lines and its natural targets are at 88, 118, and 120, so its two halves have nowhere to land.

Folding turned out to be the better shape anyway. A probe answers "can this pairing work at all" and produces evidence rather than a verdict, so it belongs on the page that also says what a graded verdict requires. Stating the boundary in one place is what stops a green probe from reading as a passed eval.

What stays gone is the test-harness-<agent> parent index for a family nobody ever authored, plus the authoring template for it.

Resolved on the branch for #1124, without an eviction. The content is back in `docs/aos-eval.md` as a closing "probe layer underneath" section: `goose-ask`'s transcript path, its `-f`/`-s`/`-m` flags, the clean-tree gotcha that makes the raw script the iteration path, and `aos-role-question`'s cloud and local mechanisms, the tmpfs argument, the ten-minute ceiling, and the `ROLE-CONFIRMED` requirement. Folded rather than given its own page for the reason above: `docs/*.md` is at the 40-doc cap, every other page is within a few lines of its own, and no pair merges under 120. The one candidate that looked splittable, `aos-roles-and-voice.md`, is 53 lines and its natural targets are at 88, 118, and 120, so its two halves have nowhere to land. Folding turned out to be the better shape anyway. A probe answers "can this pairing work at all" and produces evidence rather than a verdict, so it belongs on the page that also says what a graded verdict requires. Stating the boundary in one place is what stops a green probe from reading as a passed eval. What stays gone is the `test-harness-<agent>` parent index for a family nobody ever authored, plus the authoring template for it.

coilyco-ops referenced this issue from a commit

2026-08-18 17:17:41 +00:00

docs(aos-eval): land the probe fold and name the two consumers

coilyco-ops referenced this issue

2026-08-18 17:22:49 +00:00

docs(aos-eval): land the probe fold, name the two consumers #1128