PAN-3668

[eltmon]

Issue: #3668

Acceptance Criteria

• Add prime-agent to the canonical Harness and RuntimeName unions
• Add Prime discriminator literals to the harness-behavior unions
• Define PRIME_AGENT_BEHAVIOR and register it in the behavior map and lookup
• Add prime-agent to flywheel, artifacts, telemetry, and context-preview enumerations
• Add prime-agent to CLI --harness option help and regenerate the composer manifest
• Add the prime-agent harness literal and primeAgent section to the config schema
• Add the table-driven Overdeck-to-Prime provider and model mapping module
• Resolve the prime-agent executable with actionable missing-binary errors
• Add Prime Agent doctor checks for binary, version, RPC capability, and credentials
• Implement strict LF-delimited JSONL framing for the Prime RPC protocol
• Implement the persistent Prime RPC client: spawn, correlation, events, timeouts
• Add the Prime runtime adapter and register it in the global runtime registry
• Resume Prime sessions by verified session ID and fail loudly on mismatch
• Enforce the managed-session policy that keeps Cloister the lifecycle authority
• Launch managed work agents through the Prime RPC runtime
• Launch supervised conversations through the Prime RPC runtime
• Map Overdeck delivery, steering, abort, and kill to Prime RPC commands
• Add the Prime transcript adapter and register it in the transcript registry
• Map Prime session statistics into Overdeck cost and activity schemas
• Project Prime sessions through dashboard contracts, routes, and the read model
• Render Prime Agent in dashboard selection, badges, filters, and conversation views
• Render and inject the prime-agent-global.md context artifact
• Document Prime Agent setup, security boundary, and operational limits
• Add the mechanical Prime Agent no-loss audit
• Run full quality gates, throwaway dashboard boot, and the live Prime smoke test

[coderabbitai]

Important

Review skipped

Too many files!

This PR contains 116 files, which is 16 over the limit of 100.

To get a review, reduce the PR to 100 files or fewer by splitting it into smaller PRs or changing its base branch.

Upgrade to a paid plan to raise the limit.

This review couldn't start because sufficient usage credits or metered capacity aren't available. Add credits or update usage-based reviews in the billing tab, then retry.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: e5ee46cd-a413-40da-a4d5-182ed805dce9

📥 Commits

Reviewing files that changed from the base of the PR and between 4d551e2 and e4a4b77.

⛔ Files ignored due to path filters (2)

• docs/evidence/pan-3668/prime-agent-runtime-picker.png is excluded by !**/*.png
• packages/contracts/src/composer-commands.generated.ts is excluded by !**/*.generated.*

📒 Files selected for processing (116)

• configuration/context-layers.mdx
• configuration/harnesses.mdx
• docs/HARNESSES.md
• docs/MODEL_ROUTING.md
• docs/TELEMETRY.md
• docs/prime-agent-no-loss-audit.md
• docs/prime-agent-verification.md
• packages/contracts/src/artifacts.ts
• packages/contracts/src/context-layers.test.ts
• packages/contracts/src/context-layers.ts
• packages/contracts/src/flywheel.ts
• packages/contracts/src/harness-behavior.test.ts
• packages/contracts/src/harness-behavior.ts
• packages/contracts/src/telemetry.ts
• packages/contracts/src/types.test.ts
• packages/contracts/src/types.ts
• scripts/dashboard-types-baseline.txt
• scripts/file-size-allowlist.txt
• scripts/guard-agent-dir-removal.sh
• src/cli/commands/artifacts.ts
• src/cli/commands/doctor-prime-agent.ts
• src/cli/commands/doctor.ts
• src/cli/commands/flywheel.ts
• src/cli/commands/strike.ts
• src/cli/index.ts
• src/dashboard/frontend/src/components/CommandDeck/ConversationList.tsx
• src/dashboard/frontend/src/components/CommandDeck/ConversationRow.tsx
• src/dashboard/frontend/src/components/CommandDeck/useConversationMutations.ts
• src/dashboard/frontend/src/components/Settings/sections/CrewRow.tsx
• src/dashboard/frontend/src/components/Settings/sections/ProviderManagementSection.tsx
• src/dashboard/frontend/src/components/Settings/sections/TieredExecutionSection.tsx
• src/dashboard/frontend/src/components/Settings/types.ts
• src/dashboard/frontend/src/components/chat/ModelPicker.tsx
• src/dashboard/frontend/src/components/chat/__tests__/ModelPicker.test.tsx
• src/dashboard/frontend/src/components/sessionFeed/ConversationFeedCard.tsx
• src/dashboard/frontend/src/components/sessionFeed/useConversationFeed.ts
• src/dashboard/frontend/src/components/shared/ModelPicker/ModelPicker.tsx
• src/dashboard/frontend/src/components/shared/ModelPicker/harnessRows.ts
• src/dashboard/frontend/src/components/shared/branding/__tests__/branding.test.tsx
• src/dashboard/frontend/src/components/shared/branding/index.tsx
• src/dashboard/frontend/src/types.ts
• src/dashboard/server/routes/__tests__/conversations-supervisor.test.ts
• src/dashboard/server/routes/command-deck.ts
• src/dashboard/server/routes/context.ts
• src/dashboard/server/routes/projects.ts
• src/dashboard/server/routes/specialists/project-routes.ts
• src/dashboard/server/services/conversation-lifecycle.ts
• src/dashboard/server/services/flywheel-actions.ts
• src/dashboard/server/services/runtime-session-projection.ts
• src/lib/__tests__/harness-binary.test.ts
• src/lib/__tests__/system-prerequisites.test.ts
• src/lib/agents/delivery.ts
• src/lib/agents/provider-auth.ts
• src/lib/agents/runtime-command.ts
• src/lib/artifacts/lifecycle.ts
• src/lib/cloister/flywheel.ts
• src/lib/cloister/review-agent.ts
• src/lib/composer-commands/__tests__/harness-matrix.test.ts
• src/lib/config-yaml/__tests__/merge.test.ts
• src/lib/config-yaml/defaults.ts
• src/lib/config-yaml/merge.ts
• src/lib/config-yaml/roles.ts
• src/lib/config-yaml/schema.ts
• src/lib/context-layers/harness.ts
• src/lib/context-layers/index.ts
• src/lib/context-layers/layers.ts
• src/lib/context-layers/rendered-global.ts
• src/lib/conversations/prime-transcript-adapter.ts
• src/lib/conversations/transcript-adapter-types.ts
• src/lib/conversations/transcript-adapter.ts
• src/lib/cost-parsers/prime-agent-parser.ts
• src/lib/harness-binary.ts
• src/lib/launcher-generator.ts
• src/lib/overdeck/conversation-delivery.ts
• src/lib/overdeck/conversation-harness.ts
• src/lib/overdeck/conversation-message.ts
• src/lib/overdeck/conversation-reads.ts
• src/lib/overdeck/conversation-runtime.ts
• src/lib/overdeck/conversations.ts
• src/lib/overdeck/cost.ts
• src/lib/planning/spawn-planning-session.ts
• src/lib/prime-agent/__tests__/host-http.test.ts
• src/lib/prime-agent/__tests__/jsonl-framing.test.ts
• src/lib/prime-agent/__tests__/rpc-client.test.ts
• src/lib/prime-agent/host-http.ts
• src/lib/prime-agent/host.ts
• src/lib/prime-agent/jsonl-framing.ts
• src/lib/prime-agent/launch-command.ts
• src/lib/prime-agent/policy.ts
• src/lib/prime-agent/provider-map.test.ts
• src/lib/prime-agent/provider-map.ts
• src/lib/prime-agent/rpc-client.ts
• src/lib/prime-agent/session-controller.ts
• src/lib/prime-agent/session-resume.ts
• src/lib/runtimes/__tests__/prime-agent.test.ts
• src/lib/runtimes/index.ts
• src/lib/runtimes/prime-agent.ts
• src/lib/runtimes/types.ts
• src/lib/sync.ts
• src/lib/system-prerequisites.ts
• tests/fixtures/prime-agent/session.jsonl
• tests/integration/prime-agent-smoke.test.ts
• tests/lib/config-yaml.test.ts
• tests/lib/context-layers/harness.test.ts
• tests/unit/cli/doctor-prime-agent.test.ts
• tests/unit/dashboard/runtime-session-projection.test.ts
• tests/unit/lib/agents/spawn-prime-agent.test.ts
• tests/unit/lib/context-layers/prime-agent.test.ts
• tests/unit/lib/conversations/transcript-adapter-prime.test.ts
• tests/unit/lib/cost-parsers/prime-agent-parser.test.ts
• tests/unit/lib/overdeck/conversation-runtime-prime.test.ts
• tests/unit/lib/prime-agent-no-loss-audit.test.ts
• tests/unit/lib/prime-agent/delivery.test.ts
• tests/unit/lib/prime-agent/policy.test.ts
• tests/unit/lib/prime-agent/session-resume.test.ts
• tsdown.config.ts

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

---

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.}

[eltmon]

Review CHANGES REQUESTED for PAN-3668

Review Synthesis — PAN-3668 — 2026-08-13T07:56:50-04:00

Verdict: CHANGES REQUESTED — production Prime sessions are not connected to a persistent RPC controller

Context

• Manifest: /home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-ad27d6c5/context.json
• Branch: feature/pan-3668
• Workspace: /home/eltmon/Projects/overdeck/workspaces/feature-pan-3668
• HEAD reviewed: ad27d6c
• Cycle number: 2
• Prior cycle SHA: a85fc48

Convoy Status

Sub-role	Signal	Output	Blocking findings
security	ready	/home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-ad27d6c5/security.md	0
correctness	ready	/home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-ad27d6c5/correctness.md	2
performance	ready	/home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-ad27d6c5/performance.md	1
requirements	ready	/home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-ad27d6c5/requirements.md	4

Blocking Findings

[correctness, requirements] Production Prime sessions are not connected to the RPC controller — src/lib/launcher-generator.ts:473

The production launcher starts a bare prime-agent --mode rpc process, but no production path constructs PrimeAgentRpcClient, registers the session, or supplies a functional controller to the global runtime. Managed work agents and conversations therefore receive no initial prompt, later delivery fails against an empty registry, and lifecycle, statistics, and resume operations cannot use the promised runtime path. This deduplicates the correctness launch defect and the requirements gaps for runtime-adapter, work-agent-launch, conversation-launch, message-delivery, cost mapping, and dashboard live projection.

[correctness] Prime doctor reports RPC support without probing production binaries — src/cli/commands/doctor-prime-agent.ts:44

The --help capability probe runs only when callers inject the optional test probe. Normal production callers omit it, so any binary in the accepted version range is reported as supporting RPC even when it does not advertise --mode rpc.

[requirements] Prime doctor does not check provider credentials — src/cli/commands/doctor-prime-agent.ts:31

FR-4, FR-14, and NFR-8 require binary, version, RPC, and credential checks. The new doctor stops after binary/version/help checks, so it can declare Prime ready even when the selected provider cannot authenticate.

[requirements] Prime context preview is never rendered — src/dashboard/server/routes/context.ts:79

The response reads previews['prime-agent'], but PREVIEW_HARNESSES omits prime-agent. The API therefore returns an undefined preview instead of the effective rendered Prime context required by FR-15.

[requirements] The live smoke bypasses the production runtime path — tests/integration/prime-agent-smoke.test.ts:20

The smoke directly spawns Prime and constructs its own RPC client. It does not exercise Overdeck's production launcher, controller, registry, delivery, lifecycle, or resume entry points, so it remains green while the primary integration path is disconnected.

[performance] Dashboard list routes synchronously parse every agent transcript twice — src/dashboard/server/services/runtime-session-projection.ts:15

The new request-time projection calls token-usage and session-cost methods separately for every listed agent. Existing runtimes synchronously read and parse the complete JSONL transcript for both calls, making these endpoints block the Node event loop with work proportional to agent count times transcript size, with each transcript scanned twice.

Non-blocking Findings

[performance] Partial JSONL records are repeatedly copied and rescanned — src/lib/prime-agent/jsonl-framing.ts:29

The framer concatenates and rescans the full partial record for each chunk. An 8 MiB record split into small chunks can cause quadratic copying and scanning before the configured bound rejects or completes it; normal small records are unaffected.

Clean Sub-roles

• Security reported no findings.

Source: /home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-ad27d6c5/synthesis.md

Required action

Fix every blocking review finding, commit the fixes, then re-request review with:

pan review request PAN-3668 -m "Fixed review issues"

[eltmon]

Review CHANGES REQUESTED for PAN-3668

Review Synthesis — PAN-3668 — 2026-08-13T07:56:50-04:00

Verdict: CHANGES REQUESTED — production Prime sessions are not connected to a persistent RPC controller

Context

• Manifest: /home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-ad27d6c5/context.json
• Branch: feature/pan-3668
• Workspace: /home/eltmon/Projects/overdeck/workspaces/feature-pan-3668
• HEAD reviewed: ad27d6c
• Cycle number: 2
• Prior cycle SHA: a85fc48

Convoy Status

Sub-role	Signal	Output	Blocking findings
security	ready	/home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-ad27d6c5/security.md	0
correctness	ready	/home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-ad27d6c5/correctness.md	2
performance	ready	/home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-ad27d6c5/performance.md	1
requirements	ready	/home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-ad27d6c5/requirements.md	4

Blocking Findings

[correctness, requirements] Production Prime sessions are not connected to the RPC controller — src/lib/launcher-generator.ts:473

The production launcher starts a bare prime-agent --mode rpc process, but no production path constructs PrimeAgentRpcClient, registers the session, or supplies a functional controller to the global runtime. Managed work agents and conversations therefore receive no initial prompt, later delivery fails against an empty registry, and lifecycle, statistics, and resume operations cannot use the promised runtime path. This deduplicates the correctness launch defect and the requirements gaps for runtime-adapter, work-agent-launch, conversation-launch, message-delivery, cost mapping, and dashboard live projection.

[correctness] Prime doctor reports RPC support without probing production binaries — src/cli/commands/doctor-prime-agent.ts:44

The --help capability probe runs only when callers inject the optional test probe. Normal production callers omit it, so any binary in the accepted version range is reported as supporting RPC even when it does not advertise --mode rpc.

[requirements] Prime doctor does not check provider credentials — src/cli/commands/doctor-prime-agent.ts:31

FR-4, FR-14, and NFR-8 require binary, version, RPC, and credential checks. The new doctor stops after binary/version/help checks, so it can declare Prime ready even when the selected provider cannot authenticate.

[requirements] Prime context preview is never rendered — src/dashboard/server/routes/context.ts:79

The response reads previews['prime-agent'], but PREVIEW_HARNESSES omits prime-agent. The API therefore returns an undefined preview instead of the effective rendered Prime context required by FR-15.

[requirements] The live smoke bypasses the production runtime path — tests/integration/prime-agent-smoke.test.ts:20

The smoke directly spawns Prime and constructs its own RPC client. It does not exercise Overdeck's production launcher, controller, registry, delivery, lifecycle, or resume entry points, so it remains green while the primary integration path is disconnected.

[performance] Dashboard list routes synchronously parse every agent transcript twice — src/dashboard/server/services/runtime-session-projection.ts:15

The new request-time projection calls token-usage and session-cost methods separately for every listed agent. Existing runtimes synchronously read and parse the complete JSONL transcript for both calls, making these endpoints block the Node event loop with work proportional to agent count times transcript size, with each transcript scanned twice.

Non-blocking Findings

[performance] Partial JSONL records are repeatedly copied and rescanned — src/lib/prime-agent/jsonl-framing.ts:29

The framer concatenates and rescans the full partial record for each chunk. An 8 MiB record split into small chunks can cause quadratic copying and scanning before the configured bound rejects or completes it; normal small records are unaffected.

Clean Sub-roles

• Security reported no findings.

Source: /home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-ad27d6c5/synthesis.md

Required action

Fix every blocking review finding, commit the fixes, then re-request review with:

pan review request PAN-3668 -m "Fixed review issues"

[eltmon]

Review CHANGES REQUESTED for PAN-3668

Review Synthesis — PAN-3668 — 2026-08-13T08:57:14-04:00

Verdict: CHANGES REQUESTED — the new production host ignores the configured startup timeout and bypasses strict resume verification

Context

• Manifest: /home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-2212cbf2/context.json
• Branch: feature/pan-3668
• Workspace: /home/eltmon/Projects/overdeck/workspaces/feature-pan-3668
• HEAD reviewed: 2212cbf
• Cycle number: 3
• Prior cycle SHA: ad27d6c

Convoy Status

Sub-role	Signal	Output	Blocking findings
security	ready	/home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-2212cbf2/security.md	0
correctness	ready	/home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-2212cbf2/correctness.md	1
performance	ready	/home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-2212cbf2/performance.md	2
requirements	ready	/home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-2212cbf2/requirements.md	1

Blocking Findings

[correctness] primeAgent.rpcStartupTimeoutMs is ignored by production startup — src/lib/runtimes/prime-agent.ts:50

The new production controller hardcodes a 30-second readiness deadline, and the new host constructs PrimeAgentRpcClient with its independent 30-second default. The configured value is never passed into either boundary, so increasing or decreasing the setting has no effect. This finding is inside the cycle diff because the production controller and host were added after the prior synthesis.

[requirements] Production resume bypasses strict session identity verification — src/lib/prime-agent/host.ts:28

The new production path passes --resume directly, records whatever identity get_state returns, and never calls the existing strict verifier or compares the returned identity with the requested one. Missing or mismatched sessions can therefore be accepted as replacements, violating FR-7 and FR-9. This finding is promotable in cycle three because the cited host and launcher resume paths were added or changed after the prior synthesis.

[performance] Dashboard list routes still synchronously parse every agent transcript — src/dashboard/server/services/runtime-session-projection.ts:15

The cycle-three change consolidates two transcript parses into getSessionMetrics(), but the new implementations still synchronously locate, read, and parse one complete JSONL transcript per agent during each command-deck and project-list request. Request work remains proportional to agent count times transcript size and blocks the dashboard event loop. The cited projection and runtime metric implementations changed after the prior synthesis, so the convergence gate does not demote this unresolved blocker.

[performance] Prime host synchronously rewrites stats for every RPC event — src/lib/prime-agent/host.ts:34

The new host performs writeFileSync from the RPC event callback for every streamed event. Tool-event bursts serialize stdout processing and message-control requests behind filesystem latency, creating backpressure and delaying operator messages or aborts. The host is new in this cycle.

Non-blocking Findings

[correctness] Readiness is published before initial prompt delivery succeeds — src/lib/prime-agent/host.ts:71

The host writes its session readiness marker before sending the initial prompt. A prompt failure can therefore occur after the spawning caller has already returned success.

Clean Sub-roles

• Security reported no findings.

Source: /home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-2212cbf2/synthesis.md

Required action

Fix every blocking review finding, commit the fixes, then re-request review with:

pan review request PAN-3668 -m "Fixed review issues"

[eltmon]

Review CHANGES REQUESTED for PAN-3668

Review Synthesis — PAN-3668 — 2026-08-13T09:53:09-04:00

Verdict: CHANGES REQUESTED — failed Prime startup leaves its host, child, socket, and tmux session running

Context

• Manifest: /home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-e348f2fd/context.json
• Branch: feature/pan-3668
• Workspace: /home/eltmon/Projects/overdeck/workspaces/feature-pan-3668
• HEAD reviewed: e348f2f
• Cycle number: 4
• Prior cycle SHA: 2212cbf

Convoy Status

Sub-role	Signal	Output	Blocking findings
security	ready	/home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-e348f2fd/security.md	0
correctness	ready	/home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-e348f2fd/correctness.md	1
performance	ready	/home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-e348f2fd/performance.md	1
requirements	ready	/home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-e348f2fd/requirements.md	0

Blocking Findings

[correctness] Failed or timed-out Prime startup leaves the host and child running — src/lib/prime-agent/host.ts:117

The cycle-four lifecycle rewrite creates the child and server before initialization, but installs cleanup handlers only after initialization succeeds. Its outer catch records an error without killing those handles, and the runtime also throws on an error or readiness timeout without killing the tmux session. A failed launch can retain the provider workload, Unix socket, host, and tmux session, then race a retry for the same agent ID. This could not have been raised in cycle one because the production host did not exist then; the defective cleanup ordering was introduced by the cycle-four rewrite after the prior verdict.

[performance] Prime host buffers request bodies without a size limit — src/lib/prime-agent/host.ts:101

The changed host collects every request chunk and then allocates a second contiguous buffer for parsing, with no content-length check, streaming byte cap, or concurrency bound. Large or concurrent operator messages can exhaust the long-lived host process or cause severe garbage-collection pauses. The cited request handler was substantially rewritten in the cycle diff, so this blocker survives the convergence gate.

Non-blocking Findings

[requirements] [demoted: previously reviewed] Supervised Prime conversations always launch fresh on resume — src/lib/overdeck/conversation-runtime.ts:573

The requirements reviewer reports that the supervised-conversation branch does not pass a saved Prime session ID on resume. The cited conversation-runtime lines did not change after the prior synthesis, and the prior requirements review already identified this same missing conversation resume wiring while the prior synthesis blocked the broader strict-resume defect. Under the cycle-four convergence gate, this finding is not promotable to a blocker.

Clean Sub-roles

• Security reported no findings.

Source: /home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-e348f2fd/synthesis.md

Required action

Fix every blocking review finding, commit the fixes, then re-request review with:

pan review request PAN-3668 -m "Fixed review issues"

[eltmon]

Review CHANGES REQUESTED for PAN-3668

Review Synthesis — PAN-3668 — 2026-08-13T10:18:28-04:00

Verdict: CHANGES REQUESTED — missing Prime conversation state silently falls back to a fresh session

Context

• Manifest: /home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-b8bffc93/context.json
• Branch: feature/pan-3668
• Workspace: /home/eltmon/Projects/overdeck/workspaces/feature-pan-3668
• HEAD reviewed: b8bffc9
• Cycle number: 5
• Prior cycle SHA: e348f2f

Convoy Status

Sub-role	Signal	Output	Blocking findings
security	ready	/home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-b8bffc93/security.md	0
correctness	ready	/home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-b8bffc93/correctness.md	0
performance	ready	/home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-b8bffc93/performance.md	0
requirements	ready	/home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-b8bffc93/requirements.md	1

Blocking Findings

[requirements] Missing Prime conversation state silently becomes a fresh session — src/lib/overdeck/conversation-runtime.ts:1008

FR-7 and FR-9 require Prime resume to fail loudly when durable state is missing. The cycle-five conversation-resume changes now read the Prime session sidecar, but the outer handler still converts an absent or unreadable path into canResume = false and launches a fresh session. The helper also catches a missing sidecar and omits resumeSessionId, producing the same replacement behavior. This finding survives the convergence gate because the Prime conversation resume and sidecar-resolution paths were added or changed after the prior synthesis. It could not have been raised in cycle one because supervised Prime resume wiring did not yet exist.

Non-blocking Findings

[correctness] Prime child spawn errors bypass structured launch failure handling — src/lib/prime-agent/host.ts:86

The host does not attach a child-process error listener. If the executable becomes unavailable between resolution and spawn, the runtime eventually cleans up on timeout, but the operator receives a delayed generic readiness error instead of the real spawn failure.

Clean Sub-roles

• Security reported no findings.
• Performance reported no findings.

Source: /home/eltmon/Projects/overdeck/workspaces/feature-pan-3668/.pan/review/agent-pan-3668-review-b8bffc93/synthesis.md

Required action

Fix every blocking review finding, commit the fixes, then re-request review with:

pan review request PAN-3668 -m "Fixed review issues"