feat(desktop): merge agent edit surface into a single dialog

[wpfleger96]

Merges the two agent edit dialogs (AgentInstanceEditDialog, AgentDefinitionDialog) into a single AgentEditMergedDialog that every edit entry point — the agents library, the profile panel, owner review, and the requestOpenEditAgent event path — reaches through AgentEditDialog. AgentInstanceEditDialog is deleted, AgentDialog's instance-edit arm is removed, and the editPersonaDialogState export is gone.

Design

• One canonical form model. seedAgentFormModel builds an AgentFormModel from the edit context (linked instance, unlinked instance, or definition-only). The model owns every editable field, including the harness pin (harnessInherit / harnessCommand / harnessArgs / acpCommand). emitAgentFormDiff diffs the seed against the submitted model and routes each changed field to its owning layer, returning { personaInput, agentInput, policySets }. There is no parallel React state model and no post-emit merge.
• Single field-ownership authority. The FIELD_OWNERS map drives emit routing (emitAgentFormDiff), per-field editability (fieldEditable), and the catalog-publish dirty signal (definitionFieldsDirty). No control decides ownership locally.
• Save coordinator with observed-state settlement. runAgentSaveCoordinator applies definition → instance → local-policy writes in order. After every boundary — success or throw — it re-fetches the store and lets observed persistence decide the outcome: a write that threw after landing on disk counts as persisted and the sequence continues; a write that returned but did not persist counts as failed. A submitted behavior group (respondTo / respondToAllowlist / parallelism) settles as the full-replacement unit the backend writes, so a clear the backend failed to apply cannot false-succeed.
• Independent definition/instance runtime state. Definition-owned (dModel / dProvider) and instance-owned (iModel / iProvider) runtime state are separate, with per-owner selection handlers. An instance pin change cannot clear the definition's model or provider, and a definition runtime change cannot touch the instance overlay.
• Definition behavior fields are first-class. Access, allowlist, and parallelism render in the definition section; an unset respondTo displays as owner-only (the effective default) and is never coerced to "anyone". Blanking parallelism in a definition-only edit emits an explicit clear rather than silently restoring the stored value.
• Team-managed definitions are read-only. fieldEditable disables every definition-owned control and the dialog shows a "Managed by team «name»" notice; instance-owned fields, including access, stay editable. A runtime-less definition with a saved model but no provider exposes the provider picker and keeps Save disabled until a provider is chosen.
• Reasoning effort in one place. The EffortPicker renders in the merged dialog's instance section, so the reasoning-effort control lives on the same surface as every other editable field rather than in a separate dialog.
• Deep-link focus lands on the requested field. useAgentEditDeepLinkFocus resolves a requestOpenEditAgent focus target after the dialog mounts: it expands the collapsed env-key section when the target is a credential field, suppresses the dialog's open-autofocus so it cannot steal focus from a deep-link target, and retries once the target control is present. Focus reaches model and env_key targets that the pre-merge path left unfocused. The EnvVarsEditor focus effect keys on the required-key set's content rather than its array identity, so a content-equal re-render while model discovery and file-config queries settle no longer cancels the pending focus frame and strands the env_key target.

Out of scope (follow-up work)

• Create-flow unification
• Persona → agent vocabulary rename (profile "Instances" section, src-tauri module names)

[wolfyy970]

I do not think the failing E2E tests should be updated around this head. They are catching real lost behavior.

AgentEditDialog never calls seedAgentFormModel or emitAgentFormDiff. It submits both save inputs as null in every context, so an instance-only edit can report success without writing anything. The routed surface also removes the existing run-on, access, custom command, tuning and local policy controls. For linked agents, it writes the shared definition without showing which agents will change.

I restored the existing production routes while leaving the new model aside. The 33 affected smoke tests, the catalog publication integration, typecheck and all 4,540 unit tests pass. The flat model still needs an ownership fix too: linked respond-to, allowlist and parallelism are emitted to the instance, while definition-only changes to those fields are not emitted.

I would split this work. Keep the existing routes, then land the field model once it represents shared template state and per-agent state separately. Route one edit path at a time with parity tests. The product can still say “Edit agent” and “Edit its template”; the storage boundary does not need to become user-facing.

[wolfyy970]

Thanks for turning this around quickly. At 6f0d07f63, the follow-ups restore real instance saves and the missing run-on, access, command, tuning, environment and restart controls. The 25 focused tests, typecheck and all 4,556 unit tests pass here.

Three behavior blockers remain:

• A linked profile now always opens AgentInstanceEditDialog and explicitly removes onEditLinkedPersona. Its shared prompt, avatar and template are no longer editable from that profile. The user needs Edit this agent or Edit its template because the second choice can affect other agents.
• reviewOverrides drops request.respondTo, which the removed updateInputFromRequest preserved. A respond-to-only request opens with no requested change.
• A definition edit still shows neither the affected-agent count nor their names before save.

Please add route-level regressions for the linked-template task and the agent-requested respondTo path.

[wolfyy970]

At the current head, two concrete problems from my last pass are fixed: respond-to requests are carried through, and linked profiles can edit definition fields again. The 57 focused tests and typecheck pass here.

The remaining blocker is now clearer. In an instance-with-definition context, the form deliberately renders definition, instance and local fields together, then one Save writes all three boundaries. Changing the name, prompt or definition runtime changes the shared definition. Changing access, the harness pin or launch settings changes only this agent. The user is never asked which object they mean to edit, and linkedInstanceCount appears only in the delete copy.

Field ownership in code does not solve that user problem. These are two actions:

• Edit this agent
• Edit its template, with the affected agents shown before save

The components can be shared, but a linked definition change and an instance change should not be combined silently into one edit transaction. Please add a route-level E2E test proving the user chooses the object and sees the impact before a definition save.

[wolfyy970]

Round 8 improves field ownership, but the product blocker from my last review is unchanged. A linked agent still opens one form with definition and instance controls and one Save. There is no choice between Edit this agent and Edit its template, and the affected count and names still appear only in delete copy. The new E2E checks control visibility, not that flow.

I also found three correctness problems in definition-only editing:

• An unset respond-to policy renders as Anyone, although minting resolves unset to Owner only.
• Parallelism accepts 0, negatives, or values over 32. Save stays enabled; invalid low values are silently replaced with the old value, while high values fail in the backend.
• Returning a manually changed runtime to the auto-seeded default can drop that deliberate choice because the auto-seed marker never clears.

Please split agent and template actions, show the impacted agents before a template save, and add route-level tests for these defaults and validation.

[wolfyy970]

At 8096c37a2, the access default and runtime reselection bugs are fixed, and the R6 test now reaches the real dialog. Thanks for closing those.

Three blockers remain:

• A linked agent still opens one form where one Save can change both the agent and its shared definition. There is no choice between Edit this agent and Edit its template, and affected agents are not shown before the template write. The new E2E now locks in that combined transaction.
• Definition parallelism has no maximum or visible validation. Blank, zero and negative input is silently replaced with the saved value, while values above 32 reach backend rejection.
• If the same definition changes while this dialog is open, the form stays stale. Submit then seeds from the newer definition, combines it with stale form values and can overwrite the concurrent update. Please detect version drift or rebase the edit, with a regression test.

I am keeping changes requested at this head.

[wolfyy970]

At 8cc6e5a11, blank definition parallelism now reaches the wire as a clear, and the policy settlement change looks right. That closes the clearing defect.

Three blockers remain:

• The linked route still combines agent and template edits behind one Save. There is no object choice or affected-agent disclosure, and the E2E still requires one save to change both layers.
• Definition-only parallelism still bypasses validation. The input has no maximum, and nonblank zero or negative values silently revert to the saved value. Please cover 0, negatives, 32 and 33 through the route.
• A same-definition update arriving while the dialog is open can still be overwritten. The form stays stale, submit combines it with the latest definition seed, and the backend replaces the definition without revision checking.

I am keeping changes requested at this head.

[wolfyy970]

I re-reviewed b25ac47cf8. The observed-state save coordinator is a real improvement, but the three blockers from the last head remain.

• The linked route still puts shared definition fields and instance fields behind one Save. The E2E now explicitly requires one save to change both layers. The linked-agent count appears only in delete copy, so the user never chooses Edit this agent or Edit its template and never sees who a template edit affects.
• Definition-only parallelism still bypasses the save gate because validation runs only when an instance is present. Zero and negative values silently fall back to the saved value; 33 reaches backend rejection. Please enforce 1 through 32 in the form and cover 0, negative, 32 and 33 through the route.
• The coordinator refetches only after writing. If the definition changes while the dialog is open, it submits the stale form without an expected revision and can overwrite the newer definition. Please reject or rebase that edit before the first write.

I am keeping changes requested at this head.

[wesbillman]

Carl, an automated reviewer, commenting via Wes’s GitHub account.

Changes requested on exact head 46053a730aac1746b3265b2488d215cd32b97275, against base 00e61eafa917d296104006576b7a2ddbfd58bb5a.

The instance-name validation/emission and scoped labels/Advanced accessibility findings from the previous review are resolved. The merged-dialog product direction is unchanged. Publication recovery still has the following defects:

P1 — Combined saves bypass publication recovery

desktop/src/features/agents/ui/agentSaveCoordinator.ts:428–430 only handles publishFailed when there is no instance/policy remainder. Reproduce with a shared linked definition: change its prompt and an instance setting (or launch policy), then make strict publication preparation fail after save_personas commits but before the new event is retained (personas/update.rs:207–212; personas/pending.rs:173–200). The coordinator records publishFailed and firstError (288–294), skips instance/policy writes (299,328), and therefore bypasses the new publishRetry call. The partial warning says the profile saved and to reopen/retry (514–519), without preserving the unpublished outcome. Reopening seeds the already-saved definition; retrying only the remaining instance/policy changes emits no personaInput, so publication is never retried and this new head has no pending row to flush.

Settle publication independently of the other unsaved parts. Attempt the publish-only recovery for a persisted definition even when I/L writes remain, and report any unrecovered publication explicitly alongside those remaining changes. Add post-persist strict-preparation-failure cases with both an instance change and a policy change, including recovery success/failure. The new tests currently exercise the post-persist publication retry only in definition-only context.

P1 — New unlocked team refresh can undo an explicit unshare

desktop/src-tauri/src/commands/personas/sharing.rs:119–122 runs refresh_for_persona_at after releasing the managed-agent store lock. That helper reads the current shared team head, builds a shared=true replacement, and retains it without a transaction or a fresh visibility check (teams/pending.rs:325–365). A concrete interleaving: after a saved member prompt changes, this retry refresh reads shared head T; concurrent set_team_shared(false) acquires the store lock and retains unshared T+1; the refresh then retains its shared replacement at T+1. retain_event accepts equal timestamps (managed_agents/retention.rs:213–222), replacing the pending unshare. If the toggle's flush starts afterward, it publishes the shared replacement instead of the requested unshare. This can leave the team discoverable after the owner removed it.

Serialize the refresh's authoritative store/head reads and retain/retract operation with the same store lock as team edits/unshare/delete, without holding it across the network await. Add a controlled concurrent unshare regression at the production retry boundary. The filesystem seam's sequential pre-/post-fold tests do not cover this interleaving.

P2 — Successful publish-only retry hides skipped linked identity updates

The new success path at agentSaveCoordinator.ts:444–451,474–510 treats a successful setPersonaShared retry as complete recovery. For a shared definition name/avatar edit, strict retention failure occurs after the definition save but before linked instance name/avatar propagation and relay profile sync (desktop/src-tauri/src/commands/personas/update.rs:209–217,227–289,304–324). If the transient failure clears, the publish-only retry publishes the definition and refreshes teams (personas/sharing.rs:110–122), but never performs those skipped effects. The dialog then reports success and closes while linked identities retain the old name/avatar. A no-op reopen/save cannot recover the original name transition.

Keep local dependent effects independent of a failed enqueue, or preserve an explicit partial outcome until they are completed; do not claim full recovery from the catalog-only retry. Cover a definition-only name/avatar edit with a linked instance, a post-save strict-retain failure, and a successful publication retry, asserting the dependent identity state as well as the catalog event.

Validation: exact-head source/diff and regression-test inspection only; no checkout, build, tests, or PR-code execution. All three delegated lanes returned and were integrated. This is a bounded re-review of the agreed merged edit/save contract and changed recovery paths, not a request to restore separate dialogs or expand create-flow scope.

[wesbillman]

Carl, an automated reviewer, commenting via Wes’s GitHub account.

Changes requested on exact head 08467d5824272f10055d82df42c3f1f67f4c4e68, against base bd73490418266f267d9bb3bdf13e64582adc8e80.

The prior combined-save publication recovery and team-refresh/unshare locking defects are fixed in the inspected production paths. Linked local name/avatar persistence also now runs before a strict-retain error is returned. The merged-dialog product direction remains unchanged. Two blockers remain:

P1 — Unrelated definition edits silently erase the authored description

At agentFormModel.ts:496–505, the merged form emits a replacement definition without description. The form does not seed or expose that field either. Both update commands pass this payload through tauriPersonas.ts:113–133, where an omitted description becomes null. Rust assigns that null to the definition and saves it (personas/update.rs:218,264,283). Settlement never compares description, so the deletion is reported as a successful save.

Reproduce with an editable definition that already has a nonempty public description: open the merged editor, change only its prompt, and save. The description is cleared locally and can be published as cleared to the catalog and linked profiles. Both definition-only and linked-definition saves are affected; I/L-only saves do not issue this definition write. This is a concrete base-to-head regression: at the exact base, personaDialogState.ts:121–126 preserves the description and AgentDefinitionDialog.tsx:209–211,361–384,767–773 seeds, submits, and exposes it. This PR reroutes those edits to the merged form.

Carry description through the merged form's seed, definition ownership/diff, editing control, replacement payload, and settlement. Cover preservation on an unrelated definition edit through the real adapter for both plain save and save-and-publish, plus explicit description edit/clear behavior. Restoring a separate dialog is not requested.

P2 — Strict-retain failure still skips linked relay-profile sync

personas/update.rs:372–376 propagates retain_result? before the phase-2 sync_managed_agent_profile loop at lines 383–401. The corrective change now persists linked local records, but discards the collected kind:0 sync work on this failure path.

Reproduce with a shared definition and a linked instance with valid agent keys: edit the definition's name/avatar and make strict publication preparation fail after the definition save; let the immediate catalog-only retry succeed. set_persona_shared republishes the persona and refreshes team catalog heads, not linked kind:0 profiles (personas/sharing.rs:44–81,122–160). The coordinator then reports success and closes (agentSaveCoordinator.ts:484–550), while other clients still see the old name/avatar. Boot or a later successful start can reconcile this from the corrected local records (commands/agents.rs:1008–1032, managed_agents/restore.rs:431–485); this is stale relay metadata until reconciliation, not permanent identity loss. That later lifecycle repair does not complete the current save's skipped work.

Keep the captured retain result intact across the blocking return, release the store lock, run the existing best-effort phase-2 sync, and only then propagate the retain error. Preserve the original strict publication failure for the coordinator. Add a production-path regression with valid agent keys and a relay observer that asserts kind:0 sync is attempted even when retention fails. The new linked-instance test only checks the local name and seeds an empty private key (update/concurrent_edit_tests.rs:446,535–548), so it cannot witness this phase.

Non-blocking test notes: the new refresh observer checks the lock at its probe, not an actual concurrent unshare; and the combined-save failure test overrides updateManagedAgent without incrementing the counter it asserts (agentSaveCoordinator.test.mjs:2025–2043). Neither is an additional production blocker.

Validation: all three delegated lanes returned and were independently integrated. Exact-base/head/prior source, diffs, and regression-test inspection only. No checkout, builds, tests, or PR-code execution; no fresh CI-pass claim. This is a bounded re-review of the three agreed recovery fixes plus the concrete description merge regression, not a fresh review of unrelated incoming-main features.

[wesbillman]

Carl, an automated reviewer, commenting via Wes’s GitHub account.

Changes requested: one P2 remains

Reviewed head d739ba791332a9bb7a628593dbce64b290ecb66f against base 5aed49b505a7e27f3b0e34dafa53d6c4e8cdcd64. This is a bounded corrective re-review; the single merged editor and D → I → L save contract remain the product direction.

[P2] Include description in observed save settlement. observedStateMatchesPersonaInput never compares description. Change only an editable definition’s description to text containing U+200B. The input/adapter preserve it, and Rust rejects it before writing (personas/update.rs:218; definition_validation.rs:50–60,89–107,159–178). The unchanged definition nevertheless passes the coordinator’s per-boundary and final comparisons (agentSaveCoordinator.ts:279–288,414–423), so ordinary Save reports success and closes, discarding the draft. Save-and-publish instead treats that pre-write rejection as a publication failure, can republish the old record, and then closes successfully; combined edits can advance to I/L despite the unsaved description.

Compare description using the backend’s accepted canonical storage semantics, including absent/empty as explicit clear, without normalizing prohibited bytes into apparent success. Add production-seam regressions for rejected/nonpersisted description-only edits and clears on both ordinary and publish paths: no close, no old-record publication retry, and no I/L advancement. Also cover successful normalization and preservation on unrelated definition edits. The model, submit, and coordinator test files currently contain no description cases.

Closed in source: the prior unrelated-edit description-erasure mechanism is fixed by seed/control/emission wiring. The prior linked kind:0 sync omission is fixed by carrying the retention Result through the blocking phase, attempting profile sync outside the store lock, and only then propagating the retention error. The new native regression exercises the production publication command with valid keys and a request counter; the team-refresh test now checks retained state under an interleaving. These are inspected test oracles, not executed mutation results.

Validation: exact-tree/blob-verified source and test-text review with independent frontend/native lanes; no checkout, builds, tests, or PR-code execution. Previously closed runtime, ownership, CAS, and refetch findings were not reopened; unrelated incoming-main features and create-flow unification are excluded.

[wesbillman]

[P2] Observe description before accepting this definition as persisted

The correction emits description, but this equality helper still ignores it. For a description-only edit containing U+200B, the control/adapter preserve the text and normalize_description rejects it before any write. All fields checked here still equal the old record, so both settlement passes report persistence: ordinary Save closes as success; Save-and-publish may republish the old record and then close. Compare description with backend-compatible accepted storage semantics, including undefined-as-clear, and cover rejected edits/clears through the coordinator so they keep the dialog open and stop I/L advancement. Do not trim prohibited bytes into a false match.

[wesbillman]

Carl, an automated reviewer, commenting via Wes’s GitHub account.

Changes requested

Reviewed head a950da2a3ca2d8c669c1dc830b21e806f759f5c0 against base ac5a18697c8294e9237f505a2e01ec6fc374849a. The single edit surface is the right direction, but six user-visible regressions remain at the rendered-form/save boundaries.

1. [P2] Bind the unlinked prompt input to its editable draft

AgentEditMergedDialogInstanceSection.tsx:523–535

Open an unlinked instance, expand Advanced, and type in System prompt override. The controlled textarea receives inst.systemPrompt, but its change handler writes the dialog’s systemPrompt state. Every edit is therefore rendered back to the stored value, while buildNextAgentFormModel submits the changed state the user cannot see. Bind the control to that same draft state. Add a rendered multi-keystroke edit/Save test that checks both the visible text and submitted prompt.

2. [P2] Do not discard owner-review requests while personas are loading

AgentManagementDialogs.tsx:16–24

Delay list_personas while managed agents and channels have initialized, then deliver a valid owned-agent update request. useAgentManagement accepts it, computes matches from personasQuery.data ?? [] (119–129), and returns a not-found editError (228–238). This new effect immediately dismisses the request and retains its ID in seenRequestIds, so completion of the persona query cannot recover the draft. Distinguish loading/query failure from an authoritative missing target; do not dismiss before successful resolution. Test the actual management host with a deferred persona query.

3. [P2] Keep owner-review identity stable across a rename save

AgentManagementDialogs.tsx:52–57

In an owner-reviewed update, rename the target. matchingPersonas continually resolves the request’s original agentName against the current display name (useAgentManagement.ts:119–129). The awaited mutation invalidation and coordinator readbacks refresh that name after persistence. The match then disappears, this render guard unmounts the editor, and the dismissal effect reports “can only update … by its current name.” This can happen before save/publication settlement finishes, removing the draft/error host even when the coordinator returns a recoverable failure. Resolve the name once, retain the persona ID for the request lifetime, and test rename with delayed or failed post-write settlement. This is a source-traced interleaving, not a live reproduction.

4. [P2] Validate instance credentials against the instance overlay

useAgentEditRuntimeState.ts:332–340

Open a linked instance of a team-managed, provider-backed definition whose required API key is supplied only by the instance overlay, with no global/build/file credential. showDef selects the definition env map here; the instance required-key gate consumes that map at 363–372. The instance API-key control instead writes instanceEnvVars (AgentEditMergedDialog.tsx:889–891). Thus even a populated instance credential remains invisible to validation and Save stays disabled (575–582); the read-only definition cannot be edited to work around it. Derive the instance effective credential map from its overlay with the intended inheritance precedence, independently of D-section validation. Cover both a valid local key and an explicit-empty override.

5. [P2] Do not settle an unverified effort write as successful

agentSaveCoordinator.ts:349–358

Changing only effort produces {pubkey, effortLevel} (useAgentEditMergedSubmit.ts:287–290). If update_managed_agent fails at save_managed_agents (agent_models_update.rs:314–323) but list readback succeeds, observedStateMatchesAgentInput returns true: it checks no effort field. Both this boundary and final settlement discard the failure and close as “saved.” ManagedAgent does not expose effort, and refetchAgentStores never reads the separate config surface. Verify canonical effort before claiming success, or retain an explicitly unverified outcome. Add a production-submit regression with a rejected effort-only write and unchanged successful readback.

6. [P2] Preserve pin-to-inherit effort suppression

useAgentEditMergedSubmit.ts:284–290

For a local pinned agent with a discovered effort picker, choose a different non-null effort, select “Inherit runtime from template,” then Save. The form emits agentCommand: "" to clear the pin, but this block unconditionally appends the touched effort. The existing backend clears effort for inheritance and then applies the explicit effort field (agent_models_update.rs:93–106,197–215), restoring the override that the transition should remove. The deleted base dialog explicitly prevented this with resolveEffortSubmission({ inheritTransition: agentCommandUpdate === "" }). Restore that suppression in merged submit and test the combined picker→inherit→Save transition; a backend rewrite is unnecessary.

Scope and validation

Source-only review of immutable blobs; no PR checkout, build, test execution, or live reproduction. Carl integrated complementary UI/runtime, form/settlement, and Rust persistence/publication reviews. Coverage included library/profile/deep-link/owner-review routing, linked/unlinked/definition-only/team-managed ownership, exact-key identity, save/cancel/reopen, runtime inheritance, partial writes, CAS, and catalog publication/retry. Create-flow unification and vocabulary changes remain excluded; no independent mobile/web runtime validation was performed.

The CAS recovery candidate was rejected after tracing the awaited mutation onSettled cache invalidations; a model-clear setter mismatch was also excluded because no concrete user-impacting trigger was established. The existing effort test checks mount/hide rather than either save transition, and the R6 test mounts the merged dialog directly rather than the management host responsible for findings 2–3. Exit criteria are the six bounded fixes above with production-seam regressions, not another architectural expansion.

[wesbillman]

Carl, an automated reviewer, commenting via Wes’s GitHub account.

Changes requested: five prior contracts remain unresolved

Reviewed head d63ed1e3b96c583da3e5a9027f049b6e7cfc9af7 against base ac5a18697c8294e9237f505a2e01ec6fc374849a, bounded to the six findings from a950da2a3ca2d8c669c1dc830b21e806f759f5c0 and the new corrective delta. The single merged edit surface remains accepted product direction. Prior #6 (pin-to-inherit effort suppression) is fixed in source; #1–4 are unchanged, and #5 is incomplete at the IPC conversion boundary.

1. [P2] Bind the unlinked prompt input to its editable draft

AgentEditMergedDialogInstanceSection.tsx:523–535

Open an unlinked instance, expand Advanced, and type in System prompt override. The controlled textarea receives inst.systemPrompt, but its change handler writes the dialog’s systemPrompt state. Every edit is therefore rendered back to the stored value, while buildNextAgentFormModel submits the changed state the user cannot see. Bind the control to that same draft state. Add a rendered multi-keystroke edit/Save test that checks both the visible text and submitted prompt.

2. [P2] Do not discard owner-review requests while personas are loading

AgentManagementDialogs.tsx:16–24

Delay list_personas while managed agents and channels have initialized, then deliver a valid owned-agent update request. useAgentManagement accepts it, computes matches from personasQuery.data ?? [] (119–129), and returns a not-found editError (228–238). This new effect immediately dismisses the request and retains its ID in seenRequestIds, so completion of the persona query cannot recover the draft. Distinguish loading/query failure from an authoritative missing target; do not dismiss before successful resolution. Test the actual management host with a deferred persona query.

3. [P2] Keep owner-review identity stable across a rename save

AgentManagementDialogs.tsx:52–57

In an owner-reviewed update, rename the target. matchingPersonas continually resolves the request’s original agentName against the current display name (useAgentManagement.ts:119–129). The awaited mutation invalidation and coordinator readbacks refresh that name after persistence. The match then disappears, this render guard unmounts the editor, and the dismissal effect reports “can only update … by its current name.” This can happen before save/publication settlement finishes, removing the draft/error host even when the coordinator returns a recoverable failure. Resolve the name once, retain the persona ID for the request lifetime, and test rename with delayed or failed post-write settlement. This is a source-traced interleaving, not a live reproduction.

4. [P2] Validate instance credentials against the instance overlay

useAgentEditRuntimeState.ts:332–340

Open a linked instance of a team-managed buzz-agent/Anthropic definition whose required API key is supplied only by the instance overlay, with no global/build/file credential. showDef selects the definition env map here; the instance required-key gate consumes that map at 363–372. The instance API-key control instead writes instanceEnvVars (AgentEditMergedDialog.tsx:893–895). Thus even a populated instance credential remains invisible to validation and Save stays disabled (576–584); the read-only definition cannot be edited to work around it. Derive the instance effective credential map from its overlay with the intended inheritance precedence, independently of D-section validation. Cover both a valid local key and an explicit-empty override.

5. [P2] Carry canonical effort through the IPC mapper before comparing it

agentSaveCoordinator.ts:725–728, with the missing conversion at tauri.ts:628–671.

The Rust summary now supplies effort_level, but RawManagedAgent has no corresponding member and fromRawManagedAgent drops it. Both listManagedAgents (tauri.ts:778–781) and updateManagedAgent (1021–1030) use that mapper, so production readback still has effortLevel === undefined regardless of the persisted column. Selecting a non-null effort and successfully saving therefore compares the submitted string to null and falsely reports failure. Conversely, clearing an existing effort with a rejected write compares null to null and can close as saved while the override remains on disk. This is the still-open observed-settlement finding, with a new false-failure consequence introduced by the comparator.

Add the raw field and map it into ManagedAgent.effortLevel; test successful set and rejected clear through raw IPC responses and the real conversion/list/refetch path. The added coordinator tests construct normalized agents directly, while the rendered successful-Save test asserts the outgoing payload/count but not successful settlement, so those assertions do not catch this dropped field.

Closed finding and validation scope

Prior #6 is closed in source: useAgentEditMergedSubmit.ts:299–308 calls resolveEffortSubmission with rawAgentInput?.agentCommand === "", suppressing the explicit effort write when inheritance clears it. The runtime handler also restores the base pending-effort reset. This does not require a backend rewrite or clearing the persisted canonical effort on every runtime switch.

Source-only review of exact immutable blobs and test text. No PR checkout, build, test execution, mutation experiment, or live reproduction was performed. Re-review covered the actual unlinked prompt, linked/team-managed credential gate, owner-request loading/rename host, and effort picker → submit → native persistence → summary → IPC conversion → list/refetch settlement. Previously closed routing/ownership/CAS/catalog contracts were not reopened without new evidence. Create-flow unification, vocabulary changes, unrelated incoming-main features, and independent mobile/web runtime validation remain excluded.

Non-blocking test correction: agentEditMergedRuntimeReset.test.mjs:136,182–208 starts effort at null and changes only the touched flag before switching runtimes. Its final null assertion cannot establish that setEffortLevel(null) reset a non-null draft. Seed a real non-null selection; do not claim a mutation result from that assertion alone.

Exit criteria are the five bounded fixes above with regressions at their production boundaries, not another architectural expansion.