fix(fleet): keep adjacent dispatches reusable

[kjgbot]

Summary

• reuse recent successful control-plane evidence for adjacent team placements
• invalidate that evidence after failures and require a new roster probe after a finite lease
• delete ephemeral deterministic Relay identities during release so retries can reuse their worker names
• prepare package metadata for the workflow-only 0.1.87 next-tag publication

Why

A full fleet inventory can exceed the five-second admission boundary in a workspace with substantial historical node data. Repeating it between sequential implementer and reviewer placements can strand a half-created team. Separately, releasing only the runtime retains its create-only identity, so retrying the deterministic name collides.

The release metadata is included here so these required fixes can be deployed without merging the older release-only PR #460. Publication remains exclusively through the repository publish workflow and the next dist-tag.

Validation

• rebased onto merged PR feat(rename): migrate user-facing Factory terminology to Software Garden #466 (d780d09)
• focused fleet tests after rebase: 109 passed
• mutation proof: disabling admission reuse reproduces the five-second second-placement timeout
• mutation proof: disabling identity deletion fails the release contract assertion
• diff check passes

The complete local suite used an existing dependency tree because no install was performed: 2,468 tests passed; nine unrelated tests failed because that tree lacks the built dist artifacts and has an incompatible persona-kit export. Clean-runner CI is the merge gate.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Team

Run ID: 612e8c90-4679-4bb1-b1e6-d359b41fa7e5

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[github-actions]

@coderabbitai review

Requested for exact head be58cbd3f75303ac53e0c402aae89fe4a3326611.

[github-actions]

@coderabbitai review

Requested for exact head 0b6df4e82e89cca8864ca74093e0750ef76d6cea.

[cubic-dev-ai]

All reported issues were addressed

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[kjgbot]

Addressed all four Cubic findings on head fe4591d:

1. Failed direct roster reads now invalidate admission evidence; regression mutation fails with 2 vs 3 required roster calls.
2. Unrelated successful roster reads no longer establish or extend mutation admission evidence; the lease is now finite from the last admission probe or completed mutation. Regression mutation again demonstrates the skipped third probe.
3. Half-open mutations now perform the circuit recovery roster probe. Restoring the old pre-probe assertion makes the regression reject with FleetControlPlaneCircuitOpenError.
4. Corrected the stale commands.invoke comment to agents.release and strengthened the fake-SDK contract to record the public surface directly. Routing back through commands.invoke makes the regression observe no agent release.

Final focused suites: 112 passed; git diff --check passes. Clean-runner exact-head CI is pending.

[github-actions]

@coderabbitai review

Requested for exact head fe4591d13db5c02af2536e8d2b832c98e3595a14.

[cubic-dev-ai]

1 issue found across 4 files (changes from recent commits).

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="src/fleet/control-plane-circuit.ts">

<violation number="1" location="src/fleet/control-plane-circuit.ts:221">
P2: When `resetTimeoutMs` is shorter than `rosterTimeoutMs`, a half-open mutation can join the stale pre-open probe instead of starting recovery. That probe rejects after the generation changes, so the mutation fails and no recovery roster probe runs. Make `probe()` ignore an in-flight request from an older open generation when starting half-open recovery.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

P2: When resetTimeoutMs is shorter than rosterTimeoutMs, a half-open mutation can join the stale pre-open probe instead of starting recovery. That probe rejects after the generation changes, so the mutation fails and no recovery roster probe runs. Make probe() ignore an in-flight request from an older open generation when starting half-open recovery.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At src/fleet/control-plane-circuit.ts, line 221:

<comment>When `resetTimeoutMs` is shorter than `rosterTimeoutMs`, a half-open mutation can join the stale pre-open probe instead of starting recovery. That probe rejects after the generation changes, so the mutation fails and no recovery roster probe runs. Make `probe()` ignore an in-flight request from an older open generation when starting half-open recovery.</comment>

<file context>
@@ -204,9 +214,11 @@ export function guardFleetControlPlane(
+    // half-open must run the recovery probe even if older evidence exists.
+    const admissionState = circuit.status().state
+    if (admissionState === 'open') circuit.assertMutationAllowed()
+    if (admissionState === 'half-open' || !hasFreshAdmissionEvidence()) await probeRoster(true)
     circuit.assertMutationAllowed()
     try {
</file context>

[kjgbot]

Addressed Cubic’s follow-up P2 on head a16551a. FleetControlPlaneCircuit now tags an in-flight probe with its open generation; a half-open caller only joins a probe from the current generation and starts a fresh recovery probe when the existing request predates the open transition. Added a short-cooldown/long-roster regression. Restoring the unconditional in-flight join fails with one roster call instead of the required two; final focused suites pass 113/113 and diff check passes.

[github-actions]

@coderabbitai review

Requested for exact head a16551a98dfd3553815e711e5c73e0414ef24806.

[cubic-dev-ai]

All reported issues were addressed across 2 files (changes from recent commits).

<sub>Tip: Review your code locally with the cubic CLI to iterate faster.

Re-trigger cubic</sub>

[github-actions]

@coderabbitai review

Requested for exact head cf0698409b4c411175855b72cda6e390f3419b3c.

[kjgbot]

Addressed Cubic’s latest P2 on exact head cf06984. A probe now checks its captured open generation before recording a rejection, so a stale pre-open failure cannot reopen a circuit after the fresh half-open probe recovered it. Added a regression that opens while the old request is pending, recovers, then rejects the old request and proves the circuit remains closed. The regression failed on the prior code because the stale rejection reopened the circuit; corrected focused suites pass 187/187 and diff check passes.