skills: wait for activation completion notification (CXF-353)

[gontzess]

CXF-353 PR 2 companion (mirrors #21 for the human-activation wait).

ConductorOne c1aw now wakes the connector-authoring specialist when a human activates a minted revision (AUTHORING_JOB_COMPLETE, job_kind=ACTIVATION, payload revision + activation epoch; handler confirms ACTIVE via list_revision_summaries). This updates the skills text so the four sites agree:

• deploy-and-activate step 6: complete the S11 handoff write immediately after minting, BEFORE waiting for the activation completion notification if the host provides one; the notification is not a call, so the funnel's no-post-mint-readback discipline (S11 gate) is unchanged.
• deploy-and-activate post-activation verification: host-with-notification records ACTIVE + epoch and never polls; resume escape hatch (check list_revision_summaries once before re-minting); direct API callers poll every 5-10s, up to ~10 polls.
• author-in-app-connector human boundary: reconciled with the above (never poll ACTIVE in the funnel run; direct callers' bounded poll only in a later post-activation session).

Versions: deploy-and-activate 0.2.0 -> 0.3.0, author-in-app-connector 0.3.0 -> 0.3.1 (wording reconciliation), update-and-rollback 0.1.0 -> 0.2.0, bundle 0.5.0 -> 0.7.0, scenario/test pins updated.

Gates: npm run eval:test 139/139 pass; npm run typecheck clean.

Accepted review suggestions

• Step 6 explicitly requires the S11 handoff write immediately, BEFORE waiting for human activation; the agent ends the turn instead of withholding the handoff until the token expires.
• The fourth activation-wait site, update-and-rollback, now waits for a host notification, has the same one-time resume readback before re-minting, and reserves bounded every-5-10s polling (up to ~10 polls) for direct callers without wakeup.
• Update-and-rollback 0.2.0 and bundle 0.7.0, with manifest/scenario/test version pins aligned.
• Verification: npm run eval:test 139/139 pass; npm run typecheck:decl and npm run typecheck:parity pass.

[linear-code]

CXF-353

[github-actions]

Superseded — see the current review report for commit 53b849abd438

General PR Review: skills: wait for activation completion notification (CXF-353)

Blocking Issues: 0 | Suggestions: 3 | Threads Resolved: 0
Criteria: Criteria status: none loaded - .claude/skills/ci-review.md was not found at trusted base 4bda27274399.
Review mode: full
View review run

Review Summary

This PR only changes skill text and version pins. deploy-and-activate step 6 now tells the agent to wait for the host's activation completion notification instead of polling. Its post-activation verification section now covers three cases: a host that sends the notification, a resumed session (check once before re-minting), and direct API callers (poll every 5-10s, up to ~10 times). The author-in-app-connector human-boundary text was updated to match. Versions go from 0.2.0 to 0.3.0 (deploy-and-activate), 0.3.0 to 0.3.1 (author-in-app-connector), and 0.5.0 to 0.6.0 (bundle), and all scenario and test pins were updated to match. I scanned the full PR diff for security and correctness. The S11 safeguards are still in place: the "Do not redeem the approval token", no-force_sync and no-list_revision_summaries-during-funnel anti-patterns and locked literals are unchanged. Both skill files stay ASCII-only and under the 200-line limit. No repo-local criteria loaded, so I applied the base criteria only.

Security Issues

None found.

Correctness Issues

None found.

Suggestions

• New (medium confidence) skills/deploy-and-activate/SKILL.md:42-47: step 6 now has the agent wait for the notification after minting, but it doesn't say the S11 handoff write must happen before that wait. An agent could hold the handoff until the human activates, which can take up to the 4h token TTL. That also sits awkwardly with the "NEVER performed in the funnel run" heading on the post-activation section.
• New (high confidence) skills/update-and-rollback/SKILL.md:27-33: this skill covers the same human-activation wait but still polls list_revision_summaries unconditionally. On a host with wakeup it won't follow the new notification and resume guidance, so the PR's claim that all sites agree only holds for the funnel path.
• New (low confidence, pre-existing) evals/skills-bundle/README.md:5,73: the README still lists bundle v0.4.0 as current and pins scenario skillBundle.version at 0.4.0, so it is now two releases behind (0.6.0). Worth adding a v0.6.0 entry.

Prompt for AI agents

Verify each finding against the current code and only fix it if needed.

## Suggestions

In `skills/deploy-and-activate/SKILL.md`:
- Around line 42-47 (checklist step 6): Say explicitly that the S11 handoff write happens immediately after presenting activation_url, before any wait for the activation completion notification. Also say that recording ACTIVE + activation_epoch from the notification happens after the funnel run, so it doesn't conflict with the "NEVER performed in the funnel run" section and the "no calls after mint except the handoff write" exit criterion.

In `skills/update-and-rollback/SKILL.md`:
- Around line 27-33 (step 6): Copy the deploy-and-activate verification wording here: a host with activation completion notification delivers REVISION_STATUS_ACTIVE + activation_epoch (record them, don't poll); a resumed session with no notification checks list_revision_summaries once before re-minting; direct API callers poll every 5-10s, up to ~10 polls. Keep the locked literals "if no ACTIVE row after ~10 polls, STOP and report" and "evidence is unsatisfied". Bump the skill version and the bundle version, and update the pins in tests and scenarios.

In `evals/skills-bundle/README.md`:
- Around line 5 and 73: Add a v0.6.0 section (and v0.5.0 if it's missing) describing the notification-wait changes, and update the mount-contract pin from 0.4.0 to 0.6.0.

[github-actions]

🟡 Suggestion (medium confidence): Step 6 now says to wait for the activation notification after minting, but it doesn't say the S11 handoff write has to happen before that wait. An agent could hold the handoff until the human activates, which can take up to the 4h token TTL. The section below also says post-activation work is "NEVER performed in the funnel run". Consider saying explicitly: write the handoff first, then wait, and handle the notification-driven record of ACTIVE + epoch outside the funnel run.

[github-actions]

🟡 Suggestion (high confidence): The new notification/resume/bounded-poll guidance only went into deploy-and-activate. skills/update-and-rollback/SKILL.md:27-33 covers the same human-activation wait and still says to poll unconditionally with backoff. So on a host with wakeup, the update flow will still poll list_revision_summaries while it waits for activation. Consider mirroring this wording there too, and bumping that skill's version and the bundle.

[github-actions]

Superseded — see the current review report for commit 7804cfad1162

General PR Review: skills: wait for activation completion notification (CXF-353)

Blocking Issues: 0 | Suggestions: 3 | Threads Resolved: 0
Criteria: Criteria status: none loaded - .claude/skills/ci-review.md was not found at trusted base 4bda27274399.
Review mode: full
View review run

Review Summary

This PR only changes skill text and version pins. deploy-and-activate step 6 now tells the agent to wait for the host's activation completion notification instead of polling. Its post-activation verification section now covers three cases: a host that sends the notification, a resumed session (check once before re-minting), and direct API callers (poll every 5-10s, up to ~10 times). The author-in-app-connector human-boundary text was updated to match. Versions go from 0.2.0 to 0.3.0 (deploy-and-activate), 0.3.0 to 0.3.1 (author-in-app-connector), and 0.5.0 to 0.6.0 (bundle), and all scenario and test pins were updated to match. I scanned the full PR diff for security and correctness. The S11 safeguards are still in place: the "Do not redeem the approval token", no-force_sync and no-list_revision_summaries-during-funnel anti-patterns and locked literals are unchanged. Both skill files stay ASCII-only and under the 200-line limit. No repo-local criteria loaded, so I applied the base criteria only.

Security Issues

None found.

Correctness Issues

None found.

Suggestions

• New (medium confidence) skills/deploy-and-activate/SKILL.md:42-47: step 6 now has the agent wait for the notification after minting, but it doesn't say the S11 handoff write must happen before that wait. An agent could hold the handoff until the human activates, which can take up to the 4h token TTL. That also sits awkwardly with the "NEVER performed in the funnel run" heading on the post-activation section.
• New (high confidence) skills/update-and-rollback/SKILL.md:27-33: this skill covers the same human-activation wait but still polls list_revision_summaries unconditionally. On a host with wakeup it won't follow the new notification and resume guidance, so the PR's claim that all sites agree only holds for the funnel path.
• New (low confidence, pre-existing) evals/skills-bundle/README.md:5,73: the README still lists bundle v0.4.0 as current and pins scenario skillBundle.version at 0.4.0, so it is now two releases behind (0.6.0). Worth adding a v0.6.0 entry.

Prompt for AI agents

Verify each finding against the current code and only fix it if needed.

## Suggestions

In `skills/deploy-and-activate/SKILL.md`:
- Around line 42-47 (checklist step 6): Say explicitly that the S11 handoff write happens immediately after presenting activation_url, before any wait for the activation completion notification. Also say that recording ACTIVE + activation_epoch from the notification happens after the funnel run, so it doesn't conflict with the "NEVER performed in the funnel run" section and the "no calls after mint except the handoff write" exit criterion.

In `skills/update-and-rollback/SKILL.md`:
- Around line 27-33 (step 6): Copy the deploy-and-activate verification wording here: a host with activation completion notification delivers REVISION_STATUS_ACTIVE + activation_epoch (record them, don't poll); a resumed session with no notification checks list_revision_summaries once before re-minting; direct API callers poll every 5-10s, up to ~10 polls. Keep the locked literals "if no ACTIVE row after ~10 polls, STOP and report" and "evidence is unsatisfied". Bump the skill version and the bundle version, and update the pins in tests and scenarios.

In `evals/skills-bundle/README.md`:
- Around line 5 and 73: Add a v0.6.0 section (and v0.5.0 if it's missing) describing the notification-wait changes, and update the mount-contract pin from 0.4.0 to 0.6.0.

---

Reviewed commit: 53b849abd438

[github-actions]

No blocking issues found — see the full review report

[github-actions]

Superseded — see the current review report for commit 7804cfad1162

General PR Review: skills: wait for activation completion notification (CXF-353)

Blocking Issues: 0 | Suggestions: 1 | Threads Resolved: 0
Criteria: Criteria status: none loaded - .claude/skills/ci-review.md was not found at trusted base 4bda27274399.
Review mode: incremental since 53b849a
View review run

Review Summary

The PR changes the skill text so that, when the host provides an activation completion notification, the agent waits for it instead of polling list_revision_summaries. Callers whose host has no wakeup keep a bounded poll. The new commits change two skills. In deploy-and-activate step 6, the agent now writes the S11 handoff immediately and ends its turn before waiting. update-and-rollback step 6 gets the same notification, resume and bounded-poll guidance as deploy-and-activate (skill 0.1.0 → 0.2.0). The bundle moves to 0.7.0, and every scenario and test pin was updated to match. I scanned the full PR diff (10 files, all docs, JSON or test pins; no executable logic changed) for security and correctness and found nothing. No repo-local criteria loaded, so I applied only the base criteria. Nothing was dropped from the incremental diff and nothing was truncated.

Security Issues

None found.

Correctness Issues

None found.

Suggestions

• Prior — still present (low confidence, pre-existing) evals/skills-bundle/README.md:5,19,64: the README still calls v0.4.0 the current bundle and says the scenario skillBundle.version is 0.4.0. Every pin in bundle.json and the scenarios is now 0.7.0. This file is not in the diff, so there is no inline comment.

Resolved prior findings

• Fixed: the S11 handoff now happens before the wait. skills/deploy-and-activate/SKILL.md:42-49 says to "complete the S11 handoff write immediately, BEFORE waiting for activation. End the turn; never hold the handoff until the human activates."
• Fixed: update-and-rollback now matches the other sites. skills/update-and-rollback/SKILL.md:27-39 uses the notification first, checks once on resume, and keeps the bounded poll for direct callers, matching deploy-and-activate. The bundle pin was bumped to 0.2.0.

Prompt for AI agents

Verify each finding against the current code and only fix it if needed.

## Suggestions

In `evals/skills-bundle/README.md`:
- Around lines 5, 19 and 64: the README still calls bundle v0.4.0 current and says the scenario skillBundle.version is 0.4.0, but bundle.json and the scenarios now pin 0.7.0. Add entries for v0.5.0–v0.7.0, or at least a v0.7.0 entry: wait for the activation completion notification in deploy-and-activate 0.3.0 and update-and-rollback 0.2.0, and the wording reconciliation in author-in-app-connector 0.3.1. Change the pinned version references to 0.7.0.

[github-actions]

Superseded — see the current review report for commit 4cc16ae99439

General PR Review: skills: wait for activation completion notification (CXF-353)

Blocking Issues: 0 | Suggestions: 1 | Threads Resolved: 0
Criteria: Criteria status: none loaded - .claude/skills/ci-review.md was not found at trusted base 4bda27274399.
Review mode: incremental since 53b849a
View review run

Review Summary

The PR changes the skill text so that, when the host provides an activation completion notification, the agent waits for it instead of polling list_revision_summaries. Callers whose host has no wakeup keep a bounded poll. The new commits change two skills. In deploy-and-activate step 6, the agent now writes the S11 handoff immediately and ends its turn before waiting. update-and-rollback step 6 gets the same notification, resume and bounded-poll guidance as deploy-and-activate (skill 0.1.0 → 0.2.0). The bundle moves to 0.7.0, and every scenario and test pin was updated to match. I scanned the full PR diff (10 files, all docs, JSON or test pins; no executable logic changed) for security and correctness and found nothing. No repo-local criteria loaded, so I applied only the base criteria. Nothing was dropped from the incremental diff and nothing was truncated.

Security Issues

None found.

Correctness Issues

None found.

Suggestions

• Prior — still present (low confidence, pre-existing) evals/skills-bundle/README.md:5,19,64: the README still calls v0.4.0 the current bundle and says the scenario skillBundle.version is 0.4.0. Every pin in bundle.json and the scenarios is now 0.7.0. This file is not in the diff, so there is no inline comment.

Resolved prior findings

• Fixed: the S11 handoff now happens before the wait. skills/deploy-and-activate/SKILL.md:42-49 says to "complete the S11 handoff write immediately, BEFORE waiting for activation. End the turn; never hold the handoff until the human activates."
• Fixed: update-and-rollback now matches the other sites. skills/update-and-rollback/SKILL.md:27-39 uses the notification first, checks once on resume, and keeps the bounded poll for direct callers, matching deploy-and-activate. The bundle pin was bumped to 0.2.0.

Prompt for AI agents

Verify each finding against the current code and only fix it if needed.

## Suggestions

In `evals/skills-bundle/README.md`:
- Around lines 5, 19 and 64: the README still calls bundle v0.4.0 current and says the scenario skillBundle.version is 0.4.0, but bundle.json and the scenarios now pin 0.7.0. Add entries for v0.5.0–v0.7.0, or at least a v0.7.0 entry: wait for the activation completion notification in deploy-and-activate 0.3.0 and update-and-rollback 0.2.0, and the wording reconciliation in author-in-app-connector 0.3.1. Change the pinned version references to 0.7.0.

---

Reviewed commit: 7804cfad1162

[github-actions]

No blocking issues found — see the full review report

[github-actions]

Superseded — see the current review report for commit 4cc16ae99439

General PR Review: skills: wait for activation completion notification (CXF-353)

Blocking Issues: 0 | Suggestions: 2 | Threads Resolved: 0
Criteria: Criteria status: none loaded - .claude/skills/ci-review.md was not found at trusted base 4bda27274399.
Review mode: incremental since 7804cfa
View review run

Review Summary

The new commits simplify the activation-wait wording at all three skill sites:

• deploy-and-activate step 6 is now "present, record handoff, stop – do not wait for activation first".
• In deploy-and-activate's post-activation reference and update-and-rollback step 6, a host notification that resumes the agent now means ACTIVE, so the agent records the epoch and does not poll. On a resume with no notification, the agent checks list_revision_summaries once and then records the epoch or stops. Direct callers poll at most ~10 times, only after the OWNER reports activation.
• author-in-app-connector's Human boundary section is reconciled with this.

Skill versions are bumped (0.3.2 / 0.3.1 / 0.2.1), the bundle goes to 0.8.0, and all scenario and test pins are aligned. One locked literal changed case ("If no ACTIVE row…") and still matches the file.

I scanned the full PR diff (10 files) for security and correctness. The funnel's no-call-after-mint S11 discipline and the never-redeem / never-force_sync anti-patterns are unchanged. No repo-local criteria loaded, so only the base criteria were applied. There were no coverage gaps: the incremental diff was complete and not truncated.

Security Issues

None found.

Correctness Issues

None found.

Suggestions

• New (medium confidence) skills/deploy-and-activate/SKILL.md:54-56 (also in skills/update-and-rollback/SKILL.md step 6 and skills/author-in-app-connector/SKILL.md:118-119): the old wording tied the notification to the target revision's ACTIVE status. The new wording treats any resuming notification as "the revision is ACTIVE", with no check that the notification's revision matches the minted revision_id. That matters when several revisions exist (update/rollback).
• Prior — still present (low confidence, pre-existing) evals/skills-bundle/README.md:5,19,64: the README still calls v0.4.0 the current bundle and lists the scenario skillBundle.version as 0.4.0, but bundle.json and every scenario now pin 0.8.0. This file is not in the diff, so there is no inline comment.

Prompt for AI agents

Verify each finding against the current code and only fix it if needed.

## Suggestions

In `skills/deploy-and-activate/SKILL.md`:
- Around line 54-56: The Verification bullet says a notification resuming you "means the revision is ACTIVE". Add a requirement to confirm the notification's revision matches the target `revision_id` before recording `activation_epoch`. If it does not match, treat it as no notification for this activation.

In `skills/update-and-rollback/SKILL.md`:
- Around line 27-30 (step 6): Apply the same revision-match check to the notification branch.

In `skills/author-in-app-connector/SKILL.md`:
- Around line 118-119: Add the same check to the Human boundary wording ("the revision is ACTIVE" -> "the target revision is ACTIVE once the notification's revision matches `revision_id`").

In `evals/skills-bundle/README.md`:
- Around lines 5, 19, 64: Update the stale `v0.4.0` / `0.4.0` bundle-version references to the current `0.8.0`, and note the activation-notification changes in the version history if the README keeps one.

[github-actions]

🟡 Suggestion (medium confidence): The old text said the notification confirms the target revision's ACTIVE status. The new text says any notification that resumes the agent "means the revision is ACTIVE", and it never asks the agent to check that the notification's revision matches the minted revision_id. That gap matters in update-and-rollback, where several revisions exist. Consider adding: "confirm the notification's revision matches the target revision_id before recording activation_epoch." The same wording appears in update-and-rollback/SKILL.md step 6 and author-in-app-connector/SKILL.md (Human boundary section).

[github-actions]

General PR Review: skills: wait for activation completion notification (CXF-353)

Blocking Issues: 0 | Suggestions: 2 | Threads Resolved: 0
Criteria: Criteria status: none loaded - .claude/skills/ci-review.md was not found at trusted base 4bda27274399.
Review mode: incremental since 7804cfa
View review run

Review Summary

The new commits simplify the activation-wait wording at all three skill sites:

• deploy-and-activate step 6 is now "present, record handoff, stop – do not wait for activation first".
• In deploy-and-activate's post-activation reference and update-and-rollback step 6, a host notification that resumes the agent now means ACTIVE, so the agent records the epoch and does not poll. On a resume with no notification, the agent checks list_revision_summaries once and then records the epoch or stops. Direct callers poll at most ~10 times, only after the OWNER reports activation.
• author-in-app-connector's Human boundary section is reconciled with this.

Skill versions are bumped (0.3.2 / 0.3.1 / 0.2.1), the bundle goes to 0.8.0, and all scenario and test pins are aligned. One locked literal changed case ("If no ACTIVE row…") and still matches the file.

I scanned the full PR diff (10 files) for security and correctness. The funnel's no-call-after-mint S11 discipline and the never-redeem / never-force_sync anti-patterns are unchanged. No repo-local criteria loaded, so only the base criteria were applied. There were no coverage gaps: the incremental diff was complete and not truncated.

Security Issues

None found.

Correctness Issues

None found.

Suggestions

• New (medium confidence) skills/deploy-and-activate/SKILL.md:54-56 (also in skills/update-and-rollback/SKILL.md step 6 and skills/author-in-app-connector/SKILL.md:118-119): the old wording tied the notification to the target revision's ACTIVE status. The new wording treats any resuming notification as "the revision is ACTIVE", with no check that the notification's revision matches the minted revision_id. That matters when several revisions exist (update/rollback).
• Prior — still present (low confidence, pre-existing) evals/skills-bundle/README.md:5,19,64: the README still calls v0.4.0 the current bundle and lists the scenario skillBundle.version as 0.4.0, but bundle.json and every scenario now pin 0.8.0. This file is not in the diff, so there is no inline comment.

Prompt for AI agents

Verify each finding against the current code and only fix it if needed.

## Suggestions

In `skills/deploy-and-activate/SKILL.md`:
- Around line 54-56: The Verification bullet says a notification resuming you "means the revision is ACTIVE". Add a requirement to confirm the notification's revision matches the target `revision_id` before recording `activation_epoch`. If it does not match, treat it as no notification for this activation.

In `skills/update-and-rollback/SKILL.md`:
- Around line 27-30 (step 6): Apply the same revision-match check to the notification branch.

In `skills/author-in-app-connector/SKILL.md`:
- Around line 118-119: Add the same check to the Human boundary wording ("the revision is ACTIVE" -> "the target revision is ACTIVE once the notification's revision matches `revision_id`").

In `evals/skills-bundle/README.md`:
- Around lines 5, 19, 64: Update the stale `v0.4.0` / `0.4.0` bundle-version references to the current `0.8.0`, and note the activation-notification changes in the version history if the README keeps one.

---

Reviewed commit: 4cc16ae99439

[github-actions]

No blocking issues found — see the full review report