Fix phases index¶

Ordered execution plan for framework correctness after the security / reliability / agent-loop hunts. The stability track (fail-closed guards, toolkit factory, loop decomposition, adversarial CI gate) is largely complete -- these phases close residual security bypasses, agent-loop lies, budget gaps, and operator confusion.

Recommended execution sequence¶

Order	Plan	Effort	Risk	Depends on	Ships after merge
A	fix-phase-a-shell-containment.md	M	Med	--	Done -- restricted shell blocks `~`/`$HOME`/flag-value path bypasses; adversarial regressions in CI
B	fix-phase-b-goal-lifecycle.md	M	Med	--	Done -- Profile `max_steps` / temperature honored; `MAX_STEPS` and zombie goals handled with explicit policy + telemetry
D	fix-phase-d-budget-hard-ceiling.md	M	Med	B (spend recorded on all exit paths)	Done -- reservation hard ceiling, timeout spend accounting, optional persist, unlimited-budget clarity
F	fix-phase-f-collaboration-safety.md	M	Med	A (shared sanitization patterns)	Done -- Schedule IDOR closed; operator nudges sanitized; optional secure-profile template
C	fix-phase-c-pursuit-continuity.md	L	Med–High	B	Done -- Multi-heartbeat goals retain ReAct transcript; resume instead of amnesia each cycle
E	fix-phase-e-memory-unification.md	M	Med	C (pursuit session id stable)	Done -- One memory story: recall on pursuit + goal-gen context; `memory_set` visible to daemon semantic store
G	fix-phase-g-operator-dx-config-truth.md	M	Low–Med	D, F	Done -- Config reload contract + REST docs; toolkit cache; unified goal-gen save/budget ordering

All fix phases A–G complete. Next track: framework stabilization index (merge-gate recovery, restart continuity, security closure, operator truth, CI soak). Older deferrals remain under Out of scope (later backlog) below.

Start here: Phase A -- highest severity, smallest blast radius, no dependency on agent-loop redesign.

Parallelizable after A: B and F (independent). D should follow B so pursuit/generation spend paths are consistent before reservation logic.

Dependencies (diagram)¶

flowchart LR
  A[Phase A shell]
  B[Phase B goals]
  C[Phase C pursuit]
  D[Phase D budget]
  E[Phase E memory]
  F[Phase F collab]
  G[Phase G DX]

  A --> F
  B --> D
  B --> C
  C --> E
  D --> G
  F --> G

Effort legend¶

Size	Meaning
S	≤ 1 day, mostly tests/docs
M	2–3 days, focused module changes
L	4+ days or design + multi-PR

Risk legend¶

Level	Meaning
Low	Docs / perf-only; behavior unchanged or strictly tighter
Med	Behavior change with tests + rollback via config
Med–High	Persistence format or pursuit semantics change

Out of scope (later backlog)¶

Item	Reason
Default `guardrails.enabled: true` / `approval.enabled: true`	Product / deployment template; optional secure profile in Phase F only
Full regex guardrail rewrite	Phase F documents bypass limits; replacement engine is a separate track
PostgreSQL / second store backend	YAGNI
Full swarm routing product feature	Not framework correctness
Consolidating `hardening-guide.md` / `hardening-spec.md` / audit doc	Docs hygiene
`run_once` vs daemon `run` API parity	Phase G notes gaps; unification is API product work
Persona / life-event edge-case sweep	Phase B/G add telemetry hooks; deep persona audit deferred
HTTPS IP pinning (optional in F)	Ship only if team accepts TLS/SNI tradeoffs; HTTP pinning already exists

Relationship to stability track¶

Stability plan	Fix phase overlap
stability-03 orchestrator workspace ctor	Phase F verifies unset-workspace residual; do not duplicate if 03 merged
stability-01 budget kill-switch	Phase D adds reservation + timeout spend; builds on 01
stability-05 adversarial CI	Every phase extends `tests/adversarial/`

Merge gate command¶

uv run pytest tests/adversarial/ -v --tb=short
uv run pytest tests/ -v
uv run ruff check src/ tests/
uv run mypy src/