NEW TASK — you are a fresh Codex build agent for flower Brief #77. First run whoami() to confirm you are Solo proc 993 / project 56. Then read Solo scratchpad #1038 in this project (scratchpad_read id=1038) — your full spec + hard guardrails — and re

next steps

—

key decisions

• Fix only the read/serialization path, not write paths.
• Audit all enums surfaced in recall tools (BriefOrigin, BriefStatus, DaemonRole, DaemonStatus, HarnessType, IngestState, SegmentStatus, SessionStatus, DispatchStatus, BriefEventKind, etc.)

open questions

—

next steps

—

key decisions

• Use SafeEnumValue helper to return raw backing value when enum cast fails (ValueError) or enum is null
• Keep write paths strict so mutators still throw on unknown values
• Only patch read paths that produce MCP or UI payloads; leave write-only tools and relation logic unchanged
• Use nullsafe operator (?->value) plus fallback to raw string via helper to avoid crashes on stale DB rows

open questions

—

next steps

• Run full test suite
• Commit with Brief #82 trailer

key decisions

• Canonical scheme: flower-<role> for platform daemons, <project-slug>-<role> for project daemons
• normalize() returns null for operator:/system: refs and non-daemon refs
• Backfill uses normalize() and never touches operator:/system:
• DaemonPipelineService prefers actor_ref column when present, retains role/meta/session fallback
• No model fillable entry needed because DaemonAgent uses guarded = []

open questions

—

next steps

—

key decisions

• Reuse already-eager-loaded project relation instead of adding another query.

open questions

—

next steps

—

key decisions

• Parse JSON output in backfill dry-run test instead of relying on Laravel's line matcher for robustness.

open questions

—

next steps

—

key decisions

• Used git commit with Brief: #82 trailer.
• Staged only the files touched by Brief #82.

open questions

—

next steps

• Run `./vendor/bin/pint` to fix style
• Run full test suite (./vendor/bin/phpunit) to confirm green
• Fix any test failures or lint issues
• Commit to flower/coordination-queue with `Brief: #84` trailer
• Do NOT merge or push

key decisions

• Use a database-backed daemon_signals table for coordination instead of Solo KV/timerSet (which are unavailable in solo-cli 0.9.3)
• Dedupe identical pending signals by (kind, target_daemon_id, status=pending) to avoid stacking multiple pokes or route_feedback signals
• Retire live SoloClient methods (kvGet, kvSet, timerSet) by adding @deprecation comment and marking them as unused; the Fake SoloClient in tests throws if they are called
• Roster poke button now enqueues a 'poke' signal and shows a notice, instead of being disabled
• FeedbackPromotionService now enqueues 'route_feedback' signal for the orchestrator to pick up, removing the old routed_pending/timerSet wake path
• MCP tools (recall_signals, signal_claim/complete/fail, daemon_poke) are read-only with respect to actual Solo processes; they only manipulate daemon_signals records

open questions

—

next steps

—

key decisions

—

open questions

—

next steps

• Implement the DaemonResetState enum and migration
• Add reset_state column to daemon_agents model casts
• Create DaemonResetService with make-before-break orchestration
• Add KIND_RESET signal kind to CoordinationQueue
• Add reset action to Livewire roster UI and MCP tool
• Write tests for the reset state machine

key decisions

• Reset will reuse existing spawn/close contracts rather than inventing a parallel control path
• Baton (active orchestrator daemon id) will be stored in settings as daemon_reset.active_orchestrator_daemon_id for single source of truth
• Reset state will be its own enum/casts, separate from winddown

open questions

• Exact reset state machine transitions (states and allowed transitions)
• How the context-threshold scheduler (future Brief #86) will trigger the reset path

next steps

—

key decisions

• Baton stored as a single settings row (daemon_reset.active_orchestrator_daemon_id) for atomic compare-and-swap invariant.
• Reset is queue-only: the roster click enqueues a signal, does not directly spawn/close daemons.
• Make-before-break handoff: predecessor stays live until successor checks in and signals successor_ready.
• Service methods re-read daemon model by ID after updates to avoid stale instance state.
• MCP tools require actor_ref validation matching existing daemon tool conventions.
• Explicitly staged only the reset implementation and test files.
• Fixed return-shape of reset-retired result to override the daemon key from close result.

open questions

—

next steps

• Run the full test suite to confirm green
• Run ./vendor/bin/pint for code style
• Commit with Brief: #86 trailer on branch flower/context-threshold-reset
• Report files changed and test summary

key decisions

• Added a dedicated reset_pending boolean column instead of overloading existing reset_state
• Detection uses linked session usage as primary context source, falling back to checkin context_size
• Hold-new-work is enforced in CoordinationQueue::pending() by filtering out signals with non-reset kinds when the target daemon is reset_pending; reset lifecycle signals remain claimable
• Environment-overridable config under reset.context_threshold_tokens defaulting to 400000

open questions

—