Automaton Framework — Issues Log (from Investor Flow pilot)

Discovered while piloting the automaton framework on a real 26-slice decomposition. Each issue is categorized for the self-improvement loop. The user is concurrently developing ~/.automaton in a separate session.

P0 — Blocking: cannot dispatch implementation via pi

Issue 1: Guard blocks ALL pi-driven edits on decomposed tasks

Symptom: pi --print headless dispatch (the only pi-native way to drive a model) is blocked by the automaton-guard-pi plugin. The guard resolves the "primary task" as the decomposed parent (alphabetically first among implement-phase tasks), finds DECOMPOSITION.md, and blocks: "Task has a DECOMPOSITION.md — broken into sub-tasks for the loop runner." Even if the guard resolved to the sub-task, it finds PARENT_SPEC.md and blocks: "sub-task reserved for the loop runner." Root cause: The guard has no way to know WHICH sub-task a headless pi instance is working on. checkCanEdit calls status.py --can-edit without --task, so it picks the first implement-phase task (the parent). The guard's DECOMPOSITION.md/PARENT_SPEC.md checks then block both parent and sub-task paths. Impact: Cannot use pi --print --model <model> to implement sub-tasks. The only dispatch path that works is the loop-runner daemon via opencode run — which uses a different model config than pi. Fix direction: The guard should accept a --task argument from the pi process (e.g. via env var AUTOMATON_TASK=<sub-task-name>) so headless pi instances can declare which sub-task they're implementing. When --task points to a sub-task in implement phase, allow edits regardless of PARENT_SPEC.md.

Issue 2: Loop runner only dispatches via opencode, not pi

Symptom: loop-runner.py's _invoke_harness defaults to ["opencode", "run", "--dir", "{cwd}", "{prompt_content}"]. There is no pi-native dispatch path. Impact: Pi's configured models (omlx/Qwythos-9B, remote/ornith, etc.) are invisible to the loop runner. Opencode has its own separate model config (~/.config/opencode/opencode.json) with different provider names. To use pi-configured models, the user must either (a) mirror the config into opencode, or (b) the loop runner needs a pi dispatch mode. Fix direction: Add a pi harness option in loop.json: {"harness": {"command": ["pi", "--print", "--approve", "--model", "{model}", "{prompt_content}"]}}. Or make the harness command configurable per-loop in the template.

P1 — Phase graph friction

Issue 3: design:approved cannot transition to decomposition

Symptom: status.py --transition decomposition from design:approved fails: "Legal transitions from design:approved are: test_design, implement." The canonical flow research → decomposition → design was skipped (we went research → design), and there's no way back. Impact: DECOMPOSITION.md written during design phase is an "out-of-order artifact" that blocks transitions to test_design and implement. Required shuffling files in/out of the task folder to satisfy the gate. Fix direction: Either (a) allow design:approved → decomposition as a legal transition, or (b) allow DECOMPOSITION.md as a valid artifact in the design phase (not just decomposition), or (c) document that decomposition must always precede design and enforce it at research:approved.

Issue 4: Parent task stuck in implement after orchestrate_finalize

Symptom: automaton_orchestrate_finalize creates sub-tasks but fails to transition the parent to complete: "Failed to transition parent task." The parent stays in implement forever. Impact: Dashboard shows the parent as a permanently in-progress task. Confusing for the user. Fix direction: automaton_orchestrate_finalize should transition the parent to complete (or a new decomposed state) after sub-tasks are created.

P2 — DX / usability

Issue 5: dashboard.sh wrapper has a syntax error

Symptom: dashboard.sh uses exec python3 -c "..." without parentheses — Python 3 SyntaxError. Workaround: run PYTHONPATH=~/.automaton python3 -m automaton.dashboard directly. Fix: Change exec python3 -c to exec python3 -c → actually the issue is bash exec syntax: exec python3 -c "code" should be exec python3 -c "code" — wait, the error is SyntaxError: Missing parentheses in call to 'exec' which is a Python error inside the heredoc, not bash. The inline Python code itself has a syntax error.

Issue 6: No generic "implement-decomposition" loop template

Symptom: Only ci-triage and self-improvement loop templates ship. No template for "pick up the next new sub-task from an approved decomposition and implement it." Fix direction: Ship an implement-decomposition loop template with work_source: {kind: "decomposition"} that picks the next new sub-task, transitions it through research→implement, dispatches the implementer, then advances to code_review.

Issue 7: Slice sizing mismatch — vertical slices too large for single model dispatch

Symptom: A "vertical slice" in DECOMPOSITION.md (e.g. slice #1 = 20 files across schema/adapter/cache/queue/API/auth/UI/tests) is too large for one headless model dispatch. Both Qwythos-9B (drifted into meta-exploration) and Ornith-35B (read files but produced no code in 3 min) stalled. Fix direction: The framework should support a "micro-dispatch" layer between slices and model dispatch — each slice decomposes into 1-4 file micro-dispatches sized for the model's coherence window. Or document this as an orchestrator responsibility (which is what we did in DESIGN.md Section 8).

Issue 8: automaton_status shows project as "trader-flow" despite rename

Symptom: After renaming the repo dir trader-flow → investor-flow (with a symlink back), automaton_status still reports Project: trader-flow. The project name is cached somewhere and doesn't follow the symlink. Fix direction: Resolve the project name from the real path (resolve symlink) or from .automaton/project-name.md.

P3 — Documentation / expectations

Issue 9: No documentation that the guard blocks pi --print on decomposed tasks

The guard is designed to reserve sub-tasks for "the loop runner" — but there's no documentation that this means pi --print CANNOT be used to implement sub-tasks, and that only the opencode-based loop runner can. This caused significant confusion during the pilot.

Issue 10: automaton_orchestrate_finalize should warn about guard implications

When automaton_orchestrate_finalize creates sub-tasks with PARENT_SPEC.md, it should warn that these sub-tasks are now guard-blocked for direct pi edits and can only be worked via the loop runner.

🟠 Dead ternary in YFinanceAdapter (noted by qwopus35b backend review) — separate fix task TBD
🟡 Unused import in app/src/app/page.tsx (qwopus35b fix-spa review nit) — remove
🟡 Unused useRef import in app/src/app/page.tsx (qwopus35b fix-spa review nit) — remove
🟡 Unused Defs import in ChartLabPanel.tsx (qwopus35b re-review nit) — remove
🟡 Slice 16: add rate-limit timing test (qwopus35b approved w/ minor note) — follow-up
🟠 EdgarAdapter full_text_search does not cache ETags (qwopus found; filings_index/company_facts do) — fix adapter + tests still 1 failing
🟡 Slice 6: add tests for 13f_holdings + form4_tx (qwopus noted; bug fix approved, tests are follow-up)

7.4 KiB Raw Blame History