[SaiSatya16]

What this is

Plugin 0.2.14. Its main change: a morning check scheduled from the plugin no longer stops at its first call. Base main at f4f3efb (0.2.13, tag v0.2.13). Written against AdCopilot v2.44.0 (deployed 2026-10-22), which already has every server half. Text only: the manifest changes only its version. The tools_revision is unchanged and AdCopilot has no server change for this.

Why it ranks first. Since AdCopilot v2.42.3, Claude lists get_org_context ("AdCopilot: what is connected and where you are") under Write/delete tools. The plugin's set-up put only the read-only tools on Always allow, so every scheduled run waited at its first call for an approval nobody was there to give. A production read on 2026-10-21 found 23 scheduling workspaces whose last run fell between 10-16 and the v2.42.3 deploy, and 19 of them had not been seen since. Daily scheduled tasks are the plugin's steadiest source of real tool calls.

What is in it

Eight commits, each authored as Atromx Intelligence (the eighth is the review round below):

1. 0.2.13: a scheduled run sends claude-scheduled from every skill; the daily run still keeps no note #22, rebased onto 0.2.13 (fix/scheduled-marker @ f623fad; it was stacked on the stale 0.2.12 branch, so it is brought in as its net diff):
   • Every file that sends source: "claude-plugin/0.2.14" now also says, in one shared sentence, that a scheduled run sends source: "claude-scheduled" instead, never claude-plugin. That is 13 files besides adcopilot-daily.
   • New case review-scheduled-unmarked.
2. The daily set-up lets the run's two AdCopilot tools through:
   • skills/adcopilot-daily/SKILL.md, Claude set-up step 1, now says what AdCopilot's schedule_recipe.ALWAYS_ALLOW_STEP says (checked against origin/release/v2.44.1):
     • open Customize, then Connectors, then AdCopilot, then Tool permissions;
     • set Read-only tools to Always allow;
     • under Write/delete tools, set "AdCopilot: what is connected and where you are" and "AdCopilot: save business notes and progress" to Always allow too (they keep AdCopilot's own notes and never change Google Ads);
     • every other Write/delete tool stays on Needs approval.
   • The Team/Enterprise "Blocked" advice now leaves those two tools out.
   • The recipe handed over is v2.44.0's SCHEDULED_CHECKIN word for word. Its first call is CHECKIN_FIRST_CALL: Call get_org_context first, with source "claude-scheduled" and task "AdCopilot morning check".
   • The run's step 1 sends that task whenever the instructions open with "AdCopilot morning check", even when an older saved recipe's first call does not name it; any other scheduled task sends the task its own instructions give, or none (review round).
   • daily-scheduled-run gains a names-the-morning-check grader.
3. docs(plugin): AdCopilot v2.44's audit count: 70 checks, each listed, and the score from the checks that apply #25, squashed: v2.44's audit count. The skills now say seventy checks, each listed, with the score taken from the checks that apply.
4. test(plugin): every count of the audit's checks is AdCopilot's 70 (stacked on #25) #28, squashed (stacked on docs(plugin): AdCopilot v2.44's audit count: 70 checks, each listed, and the score from the checks that apply #25): tests/test_check_count.py.
5. 0.2.14: adcopilot-fix-checks, one yes/no a reply, the habit offer after a checked fix, one related read #24, squashed: the new adcopilot-fix-checks skill, one yes/no per reply, the habit offer only after a checked fix, and at most one related read. Its server half (adcopilot#942) is live in v2.44.0.
6. 0.2.13: an audit finding's ready call is proposed as it stands, and a draft asks for what only the customer can give #26, brought in as its net diff from 0.2.14: adcopilot-fix-checks, one yes/no a reply, the habit offer after a checked fix, one related read #24's base: an audit finding's ready call is proposed as it stands, and for a draft the skill asks for the customer's needs instead of filling them in. Its server half (A2 recommended_actions) is live in v2.44.0.
7. The README's 0.2.14 note, which covers every change above in one paragraph. The Evals section now counts thirty cases.

68 files changed (4,102 insertions, 270 deletions) against main.

Every copy of the permissions step. A grep for Needs approval / Always allow / Tool permissions finds the step in one model-read file, skills/adcopilot-daily/SKILL.md, and in the README (the "Every morning" section and the 0.2.14 note).

• commands/daily.md, the connect skill and the launch skill hand over to that set-up and carry no copy.
• AlwaysAllowStep.test_every_copy_names_both_tools fails if any paragraph in a model-read file or the README gives the step without both titles and Needs approval.
• test_the_old_step_is_gone_everywhere fails if the old wording comes back.
• The get_org_context mocks of habit-offer-chat and habit-offer-untold still quote v2.41.0's server step. They are frozen renders of that release's answer, held by the v2.41.0-fixture tests, and are not hand-edited.

Review round (2026-10-23)

Nine findings; none blocking, none on money, auth or tenant isolation. Plugin-side:

• The daily run's task could mark a customer's own schedule as the morning check (skills/adcopilot-daily/SKILL.md, the run's step 1). It sent task "AdCopilot morning check" "even when the instructions do not name it", and the run covers any scheduled task, so a customer's own task landing in this skill was counted by is_morning_check_task. Now the task goes only when the instructions open with "AdCopilot morning check" (the older saved recipe included: daily-scheduled-run's prompt is that recipe); any other scheduled task sends the task its own instructions give, or none. Held by test_the_run_names_the_morning_check_task.
• The README said nothing on AdCopilot's side stops a change in a scheduled run. Since D186 (scheduled_guard.py) the server refuses every write tool on the connection for the run (up to 20 minutes) once its first call carries claude-scheduled, AdCopilot's own notes aside. The morning section now says so, with what it does not see (a run that skips that first call or sends another source) and the Blocked advice kept. Held by test_the_schedule_section_says_what_keeps_a_run_to_reads.
• The release notes would go false once the evals ran. The 0.2.14 note and the Evals section now state run states as of when the text was written, so the paragraph the release publishes stays true whatever runs before the tag. The release script still requires "must be run before this version is tagged".
• The 0.2.13 note had been rewritten to claude-plugin/0.2.14 by 0.2.14: adcopilot-fix-checks, one yes/no a reply, the habit offer after a checked fix, one related read #24's bump. Back to 0.2.13; test_each_release_note_names_its_own_marker holds every release note to its own version.
• The squash commit on public main would carry the merging account's personal address, as f4f3efb, 8a9ab2d and the 0.2.11 and 0.2.10 squashes do. support@atromx.com belongs to the atromx-intelligence account, not the merging one, so --author-email with it cannot work; the release script now passes the merging account's GitHub noreply address, checks that gh is signed in as that account, warns if the merge commit's author differs, and has --default-author as the fallback.

The other four findings are in AdCopilot's one-off notice to people who already schedule the check, which lives outside this repository: its approval line now keeps its condition, it names both places to find the setting, a failed send no longer stops the send run, and a workspace whose morning check was seen on one day only since 2026-10-16 is now included.

Kept on purpose: the daily run still calls no save_org_context

Every schedule set up before 0.2.14 keeps that tool on Needs approval, so a run that kept a note would stop at a prompt nobody can answer. The ban lifts only after two things: AdCopilot's one-off notice to people who already schedule the check (notice_2026_10_schedule_allow, naming both tools) has been sent, and its effect has been measured. That is a later version.

Not in this release

• Plugin issue Follow-ups from v0.2.3: measure-import-trap ablation delta, mixed tool snapshot #5, the eval tool snapshots. The _tools.json files stay at the revision they were taken at, because this repository has no generator for them. Refreshing them means a whole tools/list read from the live server, never a hand edit, and that moves to 0.2.15. tests/test_check_count.py skips only the stamped snapshots.
• Nothing else slipped: 0.2.13: a scheduled run sends claude-scheduled from every skill; the daily run still keeps no note #22, 0.2.14: adcopilot-fix-checks, one yes/no a reply, the habit offer after a checked fix, one related read #24, docs(plugin): AdCopilot v2.44's audit count: 70 checks, each listed, and the score from the checks that apply #25, 0.2.13: an audit finding's ready call is proposed as it stands, and a draft asks for what only the customer can give #26 and test(plugin): every count of the audit's checks is AdCopilot's 70 (stacked on #25) #28 are all in.

Checks run (on release/0.2.14 at its head, 4e5540e)

• claude plugin validate . --strict: passed (Claude Code 2.1.309).
• claude plugin validate .claude-plugin/plugin.json --strict: passed.
• python3 -m unittest discover -s tests: 252 tests OK (251 before the review round, plus test_each_release_note_names_its_own_marker). This includes:
  • AlwaysAllowStep, which checks that every copy names both titles;
  • ScheduledSourceMarker;
  • FindingActions;
  • FixChecks;
  • CheckCount;
  • the eval-case structure tests, which cover graders, mocks and fixtures offline.
• checks.yml's other steps, run locally:
  • the manifests parse;
  • every skill and command has a name and a description;
  • no download-and-execute pattern is present;
  • the manifest (0.2.14) is ahead of the latest tag (v0.2.13).
• Not run:
  • the eval suite and the tools-only ablation (paid; the founder runs them, see below);
  • the README checklist's real-account steps 1 to 6.

Live evals (founder step, paid)

On the fixtures, never against production. The cost is about $1 a case, so roughly $30 to $35 for 30 cases x 3 runs, and about as much again for the tools-only arm. There is a $40 ceiling on each. Run it at most three times. It refuses unless production reads the version given:

bash ~/Documents/atromx/_tools/plugin-0212-evals/run.sh --live --ref 4e5540e7603c05f39aab38d02c010bb2031c11b3 --expect v2.44.1

• Pass --expect with the version that is live when you run it.
• Add --skip-ablation to run the suite alone (about $30 to $35).
• Without --live it is a dry run that writes only its plan.

Release (founder step)

The release script refuses a moved head, a moved main, a commit not authored as Atromx Intelligence or carrying a Co-authored-by trailer, an existing v0.2.14, or a folded PR merged on its own. It then marks this draft ready and squash-merges it with --match-head-commit, the squash commit authored with the merging account's GitHub noreply address rather than its personal one (--default-author is the fallback if GitHub refuses that address). It then leaves the comment "Released in 0.2.14 (#30)" on #22, #24, #25, #26 and #28 and shuts each one, makes an annotated tag v0.2.14 as Atromx Intelligence support@atromx.com, and creates the GitHub release from the README's 0.2.14 paragraph. Its --check writes nothing and passes on this head (4e5540e): bash ~/Documents/atromx/_tools/plugin-release-0214.sh --check, then the release itself is bash ~/Documents/atromx/_tools/plugin-release-0214.sh, after the evals.

Not merged. Draft until the evals have run.

🤖 Generated with Claude Code