diff --git a/.changes/work-items/fix-13875-release-check-agent-0b6e41d2.md b/.changes/work-items/fix-13875-release-check-agent-0b6e41d2.md new file mode 100644 index 00000000000..f03bb490ab4 --- /dev/null +++ b/.changes/work-items/fix-13875-release-check-agent-0b6e41d2.md @@ -0,0 +1,8 @@ +--- +bump: patch +--- + +### Changed + +- `reference/standing-item-preconditions.md`: an agent settles the `frontier-release-since-last-checked` precondition by reading the official release notes page the row's `prompt` names. With no release since `last_checked`, it runs nothing and recheck-closes the item, setting `last_checked` to the read date and advancing `next_due`. With a release, it runs the pass. Only an unreadable page leaves the item open. +- `scripts/evaluate-schedule-precondition.sh` prints the row's `prompt` ahead of `met` when the precondition has no operator gate. A `work` or `recheck` caller then gets the steps it must settle, instead of a bare `met` that reads as already checked. diff --git a/.github/recurring-schedule.json b/.github/recurring-schedule.json index d445d3cd79e..45c89908455 100644 --- a/.github/recurring-schedule.json +++ b/.github/recurring-schedule.json @@ -9,12 +9,12 @@ "triggers": ["claude-model-release"], "precondition": { "id": "frontier-release-since-last-checked", - "prompt": "A Claude model release of any class (frontier, Opus, Sonnet or Haiku) must have occurred since last_checked. audit-instructions runs on any release; unhobble runs only when the release is a frontier model. If no release has occurred, run nothing and leave last_checked and next_due unchanged; never recheck on cadence alone.", - "requires_operator_confirmation": true + "prompt": "Read the official Claude release notes (https://platform.claude.com/docs/en/release-notes/overview) for a Claude model release of any class (frontier, Opus, Sonnet or Haiku) after last_checked. audit-instructions runs on any release; unhobble runs only when the release is a frontier model. If no release has occurred, run nothing and recheck: cite the page and its newest model entry in the recheck, set last_checked to the date you read it, and advance next_due by the cadence. If the page cannot be read, run nothing and leave the issue open.", + "requires_operator_confirmation": false }, "last_checked": "2026-10-10", "next_due": "2026-11-06", - "notes": "On each Claude model release (primary trigger), re-test which standing instructions the new model still needs over this repository's session surfaces (CLAUDE.md, AGENTS.md, .claude/rules, local hooks and settings), never the shipped plugin components. /harness-config:audit-instructions runs report-only on any release (frontier, Opus, Sonnet or Haiku). On a frontier release only, /harness-config:unhobble follows on the survivors: strip candidate instructions on an experiment branch, exercise the affected workflows, restore only what observed failures justify, and run /overengineering:delta as a step of that pass, so each frontier release re-checks whether the enforcement surface (guards, hooks, gates) still earns its keep. Record the audit's findings summary and, for unhobble, what was removed, what was restored with the evidence for each, and what the delta moved, in the body of the pull request that advances last_checked (/work-items:track recheck). Shipped-component findings go to their own issues. The id names only the frontier case; the precondition prompt sets the release class. No automation observes the release event: /work-items:track due lists this row once next_due arrives, and an operator or agent who notices a release runs it sooner. The quarterly cadence is a backstop so the pass happens even when a release goes unnoticed (docs/plugin-philosophy.md, Instruction economy). 2026-10-10: Claude Haiku 5.5 (not a frontier release) got the audit-instructions pass only; its findings are on #2019.", + "notes": "On each Claude model release (primary trigger), re-test which standing instructions the new model still needs over this repository's session surfaces (CLAUDE.md, AGENTS.md, .claude/rules, local hooks and settings), never the shipped plugin components. /harness-config:audit-instructions runs report-only on any release (frontier, Opus, Sonnet or Haiku). On a frontier release only, /harness-config:unhobble follows on the survivors: strip candidate instructions on an experiment branch, exercise the affected workflows, restore only what observed failures justify, and run /overengineering:delta as a step of that pass, so each frontier release re-checks whether the enforcement surface (guards, hooks, gates) still earns its keep. Record the audit's findings summary and, for unhobble, what was removed, what was restored with the evidence for each, and what the delta moved, in the body of the pull request that advances last_checked (/work-items:track recheck). Shipped-component findings go to their own issues. The id names only the frontier case; the precondition prompt sets the release class. No automation observes the release event: the daily recurring filer files this row once next_due arrives, so the quarterly cadence is a backstop that catches a release nobody noticed (docs/plugin-philosophy.md, Instruction economy), and an operator or agent who notices a release runs it sooner. 2026-10-10: Claude Haiku 5.5 (not a frontier release) got the audit-instructions pass only; its findings are on #2019.", "close_previous": true }, { diff --git a/plugins/work-items/reference/standing-item-preconditions.md b/plugins/work-items/reference/standing-item-preconditions.md index e0aba0d9c2d..8be511b04e2 100644 --- a/plugins/work-items/reference/standing-item-preconditions.md +++ b/plugins/work-items/reference/standing-item-preconditions.md @@ -27,19 +27,22 @@ renaming it would break the schedule rows and scripts that match it. Before claiming (tier 4) or recheck-closing: 1. Read the row's `last_checked` date. -2. Ask whether a Claude model release of the class the row's `prompt` names occurred - **after** that date. -3. If **no** (or unknown): **do not claim, do not recheck-close**. Leave the open - recurring issue open and report the `prompt` text inline. -4. If **yes**: proceed, and record in the claim/recheck comment that the operator - confirmed a post-`last_checked` release. +2. Read the official release notes page the row's `prompt` names for a Claude model + release of the class it names **after** that date. +3. If **no**: run nothing and recheck-close. Cite the page and its newest model entry, + set `last_checked` to the date you read it, and advance `next_due` by the cadence. +4. If **yes**: run the pass, and cite the release in the claim and recheck comments. +5. If the page cannot be read: do not claim and do not recheck-close. Leave the + issue open and report why. -There is no automated release oracle in-repo; `requires_operator_confirmation: true` -makes the confirmation explicit rather than inferred from cadence. +The release notes page is the release oracle, so an agent can settle the check. A row +sets `requires_operator_confirmation: true` only when a person must confirm it. ## Seam helper `plugins/work-items/scripts/evaluate-schedule-precondition.sh` prints `met`, `unmet`, or `needs-confirmation` for a schedule row id. Pass `--operator-confirmed` when the operator has affirmed a confirmation-style -precondition. +precondition. For a row with no operator gate, `met` means only that nothing +holds the row: the script prints the row's `prompt` ahead of it, and the caller +settles the check with the `id` steps above before running the pass. diff --git a/plugins/work-items/scripts/evaluate-schedule-precondition.sh b/plugins/work-items/scripts/evaluate-schedule-precondition.sh index 8eaf272fdf3..0f12f69371e 100755 --- a/plugins/work-items/scripts/evaluate-schedule-precondition.sh +++ b/plugins/work-items/scripts/evaluate-schedule-precondition.sh @@ -60,6 +60,12 @@ frontier-release-since-last-checked) echo "needs-confirmation" exit 2 fi + # With no operator gate the caller settles the check itself, so it gets the + # prompt's steps (standing-item-preconditions.md) ahead of the marker. + if [[ "$requires_confirm" != "true" ]]; then + prompt="$(jq -r '.precondition.prompt // empty' <<<"$row")" + [[ -z "$prompt" ]] || printf '%s\n' "$prompt" + fi echo "met" exit 0 ;; diff --git a/plugins/work-items/scripts/evaluate-schedule-precondition.test.sh b/plugins/work-items/scripts/evaluate-schedule-precondition.test.sh index 0c6ed465e5a..529cfc8fcae 100755 --- a/plugins/work-items/scripts/evaluate-schedule-precondition.test.sh +++ b/plugins/work-items/scripts/evaluate-schedule-precondition.test.sh @@ -18,6 +18,15 @@ cat >"$TMP" <<'JSON' "requires_operator_confirmation": true } }, + { + "id": "agent", + "title": "Agent-checked", + "precondition": { + "id": "frontier-release-since-last-checked", + "prompt": "read the release notes", + "requires_operator_confirmation": false + } + }, { "id": "plain", "title": "Plain" @@ -64,6 +73,14 @@ else fail "confirmed path" fi +out="$("$EVAL" "$TMP" agent)" +rc=$? +if [[ "$rc" -eq 0 && "$out" == $'read the release notes\nmet' ]]; then + pass "a precondition with no operator gate prints its prompt, then met" +else + fail "agent-checked path exit=$rc stdout: $out" +fi + if [[ "$FAILED" -eq 0 ]]; then exit 0 fi