Skip to content

HANDOFF.md

Latest commit

 

History

History
132 lines (116 loc) · 9.08 KB

HANDOFF.md

File metadata and controls

132 lines (116 loc) · 9.08 KB
Raw

Session Handoff — 2026-09-05

Where things stand

Symphony (oh-my-symphony) gained its dark-factory cycle on branch feat/dark-factory-cycle (from dev): a software request typed into the web chat becomes a strict intent card, the operator approves once, and the server files the request ticket into the pipeline. That approval is the only human gate; the deep preset does the rest. Design and plan live in docs/plans/2026-09-05-dark-factory-cycle-design.md and docs/plans/2026-09-05-dark-factory-cycle-implementation-plan.md.

Every gate is green on the branch:

  • ruff ✓ · pyright 0 errors ✓ · i18n 566 keys (en/ko) ✓
  • Full pytest suite 2476 passed, 14 skipped (baseline before the branch: 2415 passed, 14 skipped)
  • Live run-path check: symphony service start on a scratch project, a real claude chat session in Q&A mode produced the intent card from one request, the bare approve reply filed REQ-1 (request: python-todo-cli) in Todo and wrote .sdlc/work/python-todo-cli/intent.md with its ## Approval section.

What the branch did

Area Content
Intent gate src/symphony/intent.py (marker parser, IntentAction, file_intent_request), chat.py (confirm_intent, intent_for_reply, supersede/expiry/prune, agent notice), webapi.py (POST /api/v1/chat/sessions/{sid}/intent/{action_id}/approve, approve reply routing), web card (app.js, i18n.js, style.css)
Pipeline _BOARD_PREAMBLE rewritten around _INTENT_PROTOCOL; deep intake.md consumes the intent (Track: micro → Plan); project-setup proposals accept preset: deep and projects.create_or_adopt_project(preset=) applies it before the initial commit
Bugs #29 per-dispatch env in BackendInit.env; #28 RetryEntry.touched_files snapshot for the conflict pre-check; #30 utils/atomic_json.py rename retry; #32 per-workflow state file names
Debt orchestrator/worker_exit.py (worker-exit state machine, 1009 lines) and orchestrator/attempt.py (_run_agent_attempt phases, 1095 lines) extracted from core.py (11349 → 9815 lines); claude and pi backends now subclass per_turn.JsonlStreamBackend (claude_code.py 503 → 267, pi.py 666 → 386)
Docs README (en/ko) "Chat intake" section, CONTEXT.md (Intent proposal, Intent approval, Track), docs/PIPELINE.md intent gate, skills/symphony-skill/reference/operations.md, docs/architecture.md module table, CHANGELOG

End-to-end run (2026-09-05, scratch project, real agents)

Scratch registry: SYMPHONY_PROJECTS_FILE=/tmp/symphony-intent-smoke.ntna/projects.json; the deep project lives at /tmp/symphony-intent-smoke.ntna/todo-app (service port 9999, log log/symphony.log, stop with symphony service stop ./WORKFLOW.md from that directory). The pipeline was still finishing when this handoff was written; check the board there.

Step What happened Evidence
Deep project from chat On the default smoke board, an edit-mode codex chat proposed a separate project with preset: deep; one confirmation created and registered todo-cli with the eight lanes e2e-step2-codex.log, projects.json
Intent gate On the new board, a Q&A codex chat produced the intent card from one request; a bare approve reply filed REQ-1 in Intake with request: python-cli-todo and wrote .sdlc/work/python-cli-todo/intent.md e2e-step3.log, kanban/REQ-1.md
Request lanes REQ-1 walked Intake → Research → Plan → Review → Done with no operator question; Review passed; the branch merged (auto_merge_completed) log/symphony.log
Decomposition Plan spawned BUILD-1..3, QA-1, VERIFY-1, DOCUMENT-1 and wrote docs/req/python-cli-todo/{brief,research,plan,contracts,review}.md plus release-contract.yaml scratch main
Builds BUILD-1, BUILD-2, BUILD-3 each reached Done and merged in order; QA-1 reached Done at 23:10 UTC; VERIFY-1 and DOCUMENT-1 were still pending monitor log
Delivered app git archive main of the scratch project: todo.py ($TODO_FILE override, JSON next to the script), tests/test_{storage,cli,readme}.py, README; pytest -q → 26 passed; todo.py add "buy milk" then list prints [ ] 1. buy milk export run at 23:09 UTC

Findings from the run (not code defects in Symphony unless noted):

  • The claude CLI hit the account's session limit; the chat turn failed cleanly (turn_failed in the transcript) and no card was produced.
  • codex hit its usage limit mid-Plan; the worker was auto-paused (worker_error_auto_paused). Recovery that worked: edit agent.kind in WORKFLOW.md and the ticket's own agent.kind (the file tracker stamps the dispatched kind on the ticket and it overrides the board default), then POST /api/v1/REQ-1/resume. WORKFLOW.md edits are picked up live; no restart was needed.
  • ~/.opencode/bin/opencode on this host is a symlink to a broken npm postinstall stub; the real 1.18.18 binary is ~/.opencode/bin/opencode1. The scratch workflow uses that absolute path in opencode.command. Product follow-up worth considering: symphony doctor could run <agent> --version for the configured kind and flag a non-zero exit.
  • Wall clock: intent approval 22:29 UTC → REQ-1 Done 22:59 → three builds merged by 23:07 → QA Done 23:10, with ~20 minutes lost to the two quota failures.

Open questions / next steps

  1. Deep boards from the example workflow cannot dispatch the app-release verifier (found in the E2E, 23:11 UTC). WORKFLOW.file.example.md ships feature_base_branch: "" and auto_merge_target_branch: "" (meaning "the current branch"); symphony doctor accepts that for the deep preset ('' == ''), but orchestrator/release_contracts.py refuses to bind the verifier: release_dispatch_refused ... configured target_branch must name a resolvable local branch: '', every tick, while VERIFY-1 sits in Verify and the finalizer waits on it. Setting both keys to "main" in the scratch workflow (live reload) dispatched VERIFY-1 within a tick. Fix candidates: resolve an empty configured target to the repository's current branch inside the binder (matching auto-merge semantics), or have the preset: deep bootstrap write the repository's initial branch into both keys; either way check_deep_preset_merge_contract and the binder must agree. Add a regression test that binds a verifier on a board whose configured target is empty.

  2. #31 codex sandbox vs symlinked board files. Current code already injects resolved symlink targets and git admin dirs (backends/codex.py:_scan_workspace_symlinks). The 2026-05 failure has no deterministic repro; it needs a live codex run against a symlinked file board before it can be closed or fixed.

  3. Next core.py extractions (in order): _enforce_app_release_transition_inner (~420 lines) into release_cycle.py; _on_tick (~400); _reconcile_one (~345). Follow the worker_exit.py convention: module-level functions that take orch explicitly, _core() late binding for names tests patch on core, full suite as the gate.

  4. win32 process_identity() still returns None (kills stay ungated warn-once on Windows). A GetProcessTimes-based fingerprint in _shell.py would enable pid-reuse protection.

  5. Deferred UX items from the 2026-08-23 audit: scrollable HelpScreen, modal focus trap, web-board rendering for Linear/Jira trackers, TUI↔web action parity.

  6. Intent gate follow-ups: Linear/Jira boards are not covered (proposals parse only on file boards); the kit's tools/tripwire.sh hedging scan is not run on the intent body; this repository's own .sdlc/config.md still has no lazymode: line (default 1) — set lazymode: 3 only on the owner's explicit instruction.

  7. tests/test_orchestrator_dispatch.py::test_startup_reclaim_terminates_live_recorded_orphan_agent_group is timing-flaky under CPU load (it kills a real process group). It passed in every full-suite run on this branch; rerun it alone before treating a failure as a regression.

  8. Parallel test suite (pytest-xdist) is not isolation-clean yet. pytest -n auto finishes in ~72 s instead of ~234 s, but two tests fail under parallel load and pass serially: tests/test_workspace.py::test_setup_worktree_script_supports_linked_workflow_dir[mkdir] and tests/test_orchestrator_release_contract_integration.py::test_reconcile_terminal_release_holds_lease_until_cleanup_finishes (an asyncio timeout). Make those two load-tolerant before adding pytest-xdist to the dev extra and -n auto to CI; nothing in the repo depends on xdist today.

Environment notes

  • Run tests with .venv/bin/python -m pytest -q (bare python is not the project interpreter).
  • symphony is not on the login-shell PATH on this host; workers receive SYMPHONY_CLI=<venv>/bin/symphony, and the shipped prompts use ${SYMPHONY_CLI:-symphony}.
  • Scratch project used for the live check: created with SYMPHONY_PROJECTS_FILE=<tmp>/projects.json symphony project create ... --port 9997 so the global project registry stayed untouched.