RF — TFW-52 / Phase B: Assisted¶
Date: 2026-08-13 Author: Codex (Executor) Status: 🟢 RF — Complete Owner disposition: SUCCESS WITH KNOWN LIMITATION Parent HL: HL-TFW-52 TS: TS Phase B
1. What Was Done¶
New Files¶
| File | Description |
|---|---|
editions/02-assisted/README.md |
Standalone Russian Assisted guide: initialization, working cycle, manual fallback and honest boundaries. |
editions/02-assisted/AGENTS.md |
Agent contract for project initialization, critical TFW mode, task discovery, traces, memory, risk and migration. |
editions/02-assisted/PROJECT.md |
Uninitialized standalone project card for goals, user, audience, success criteria, values, participants and constraints. |
editions/02-assisted/MIGRATION.md |
Preserving Light → Assisted migration procedure with duplicate-contract stop. |
editions/02-assisted/.codex/hooks.json |
Project lifecycle definitions for SessionStart, PreCompact and Stop. |
editions/02-assisted/.codex/hooks/tfw-hook.ps1 |
Windows handler for root, actor, state, checkpoints, consistency and deterministic secret checks. |
editions/02-assisted/.codex/hooks/tfw-hook.sh |
POSIX counterpart of the handler. |
editions/02-assisted/people/README.md |
Separate human/automation profile and declared-attribution contract. |
editions/02-assisted/knowledge/INDEX.md |
Candidate/record navigation with explicit non-automatic semantics. |
phase-b/evidence/EV__phase-b__assisted.md |
Structured evidence ledger, lifecycle blocker and owner disposition. |
phase-b/RF__phase-b__assisted.md |
This execution result. |
Modified Files¶
| File | Changes |
|---|---|
editions/README.md |
Marks Assisted as available and links the standalone starter. |
README.md |
Adds Assisted to the editions entry and transitions Phase B to RF. |
phase-b/ONB__phase-b__assisted.md |
Records approved onboarding and expected evidence dispatch plan. |
Implementation commits: 8557a32 (initial Assisted starter), ac4c3a4 (owner-directed standalone initialization and active task intake). ONB commit: 5f4f308. No push was performed.
Scope accounting: 14 touched repository files in Phase B, 12 new including ONB/EV/RF, 2 modified, 844 product LOC; within 30 files / 15 new / 3000 LOC / 30 modified.
2. Key Decisions¶
- Assisted is a standalone edition. It no longer assumes Light context or a pre-filled generic goal.
PROJECT.mdstarts uninitialized and the first-run dialogue captures the actual project, user, audience, goal, desired effect, success criteria, values and constraints. - The user’s preferred professional role and mental model belong in the active agent contract. Initialization fills the two designated values in
AGENTS.md, then requires a new session so the changed contract is loaded. - TFW’s critical mode is explicit. The agent must be useful rather than agreeable: no flattery, evidence-backed disagreement, facts separated from conclusions, unknowns named, Working Backwards before action and no placeholders.
- Creating a task is not useful progress by itself. Vague problems trigger up to three high-value discovery questions. After creating a trace, the agent must immediately take the next useful step or ask a concrete blocking question.
- Hooks remain delivered but not claimed as working lifecycle evidence. Deterministic handlers passed local fixtures; trusted Codex Desktop did not durably dispatch them in the tested runtimes. The manual fallback is therefore a real operating mode, not hidden degradation.
- Owner acceptance governs product closure, not evidence history. The owner directed successful closure after basic use. EV preserves
AC-2 BLOCKEDand unexecuted lanes asDEFERRED; success does not rewrite observations.
3. Acceptance Criteria¶
- [x] AC-1 — Assisted opens as a standalone project. The nine-file starter copies to a clean root, explains limits and initialization, contains no user-facing JSON/code requirement, and respects the 700/1100 word limits.
- [ ] AC-2 — Hooks installed, trusted and executed. BLOCKED. Definitions and handlers exist and pass deterministic fixtures, but supported Codex Desktop did not produce durable lifecycle dispatch evidence.
- [ ] AC-3 — Durable action does not start without a trace. DEFERRED live gate. The final contract requires orientation and trace-before-result; real Stop mismatch evidence was not completed.
- [ ] AC-4 — Ordinary conversation creates nothing. DEFERRED full lane. Contract and an incomplete safe no-write run support the behavior; the complete three-question and
active_task=nonelifecycle proof was not collected. - [x] AC-5 — Root is resolved without Git. Root/nested/missing/competing-marker fixtures passed and the final uninitialized project marker remains discoverable.
- [ ] AC-6 — Memory is prepared but not presented as automatic. DEFERRED live candidate gate. Structure and wording pass inspection; AC-8 candidate artifacts were not collected.
- [ ] AC-7 — Risk gate precedes shared writing. DEFERRED full lane. Deterministic secret fixture and semantic hold contract exist; the two full live runs and Stop publication check were not completed.
- [ ] AC-8 — Light/Assisted comparison. DEFERRED. The frozen scenario was not run; no measured-difference claim is made.
- [ ] AC-9 — Light migration preserves the project. DEFERRED. The preservation procedure exists; live migration and byte comparison were not run.
- [ ] AC-10 — Two participants share a non-Git root. DEFERRED. Ownership/profile design exists; parallel live evidence was not run.
- [x] AC-11 — Visible manual mode. Exact fallback wording and manual order are documented and were observed in the incomplete interactive run; this does not verify lifecycle hooks.
Owner outcome: accepted as successful with the hook limitation and deferred evidence explicitly recorded.
4. Verification¶
- Structural assertions: PASS — standalone wording, initialization prompt/state, project goals and values, agent role/mental model, critical mode, task intake and no-stop-after-trace checks all passed.
- Word limits: PASS —
README.md531/700;AGENTS.md1038/1100. - Handler fixtures: PASS — root/nested/missing/competing-root paths, trace consistency, checkpoint idempotence, one-shot Stop state and deterministic secret checks were exercised locally.
- Fresh-root smoke after the final patch: PASS — final
PROJECT.mdwas resolved from a nested directory in a non-Git root;SessionStarthandler returned a valid summary and created neitherwork/norknowledge/inbox/. - Git diff checks: PASS — product commits contain only owned Phase B paths; no
.tfw/, TFW-51, substantive Light, TFW-53 or TFW-54 product changes. - Codex lifecycle: BLOCKED — trusted interactive Codex sessions did not yield durable dispatch events, handler state or additional context. No synthetic call is promoted to lifecycle evidence.
- Full downstream live suite: DEFERRED by owner closure — comparison, migration, two-participant and remaining risk lanes were not rerun after final owner corrections.
5. Evidence¶
See EV file for evidence details.
Evidence verdict: 3/11 VERIFIED, 7 DEFERRED, 1 BLOCKED, 0 N/A
Product disposition: SUCCESS WITH KNOWN LIMITATION by explicit owner direction on 2026-08-13.
6. Observations (out-of-scope, not modified)¶
| # | File | Line(s) | Type | Description |
|---|---|---|---|---|
| 1 | Codex Desktop lifecycle runtime | N/A | missing-test | Project hooks were visible/trusted but not durably dispatched in tested Desktop runtimes. Upstream runtime correction is required before AC-2 can be reverified; global hooks, trust bypass and product redesign were intentionally not used. |
| 2 | External raw evidence roots | N/A | ux | The previously reported external run roots and rollout files were absent at RF closure on 2026-08-13. EV retains exact session IDs and outcomes, but raw reviewer-inspectable bundles are unavailable. |
7. Fact Candidates¶
| # | Category | Candidate | Source | Confidence |
|---|---|---|---|---|
| 1 | product | Assisted must be independently usable and must not assume Light is installed alongside it or was used as an upgrade source. | Owner feedback, 2026-08-09 | High |
| 2 | process | At initialization, the method must capture the actual project, user, goal, desired effect, success criteria and values, plus the professional role and mental model expected from the AI. | Owner feedback, 2026-08-09 | High |
| 3 | philosophy | The baseline TFW agent mode is a critically useful partner: it challenges weak assumptions with evidence, does not flatter or agree automatically, and remains honest about uncertainty. | Owner feedback, 2026-08-09 | High |
| 4 | ux | A vague request must lead to problem discovery and useful follow-through; merely creating a task and waiting is an unacceptable assistant behavior. | Owner garden/gardener test, 2026-08-09 | High |
| 5 | product | The owner accepts the current Assisted edition as successful for closure while treating broken lifecycle hooks as a known external limitation. | Owner closure direction, 2026-08-13 | High |
8. Strategic Insights (Execution)¶
| # | Insight | Category | Source |
|---|---|---|---|
| S1 | Structure can enforce trace hygiene yet still produce an inert assistant if discovery and forward motion are underspecified. Implication: every edition must preserve both halves of the method: durable traces and an active, critical problem-solving stance. | philosophy | Owner’s standalone and garden/gardener feedback |
| S2 | Standalone initialization is product behavior, not setup documentation. Implication: a starter’s initial files must visibly represent “unknown project” and instruct the agent to construct project-specific context through dialogue instead of shipping a plausible generic goal. | product | Owner’s basic Assisted test |
| S3 | Owner-level product acceptance and formal evidence completeness answer different questions. Implication: TFW can close a useful MVP successfully while retaining precise blocked/deferred evidence, provided no unverified mechanism is relabeled as working. | process | Owner closure direction after hook failure |
9. Diagrams¶
flowchart LR
U["First user message"] --> I{"Project initialized?"}
I -- "No" --> Q["≤3 questions: project, goal/values, AI role/model"]
Q --> P["PROJECT + profiles + AGENTS settings"]
P --> N["New session"]
I -- "Yes" --> D{"Request clear enough?"}
N --> D
D -- "No" --> C["≤3 high-value discovery questions"]
D -- "Yes" --> T["TRACE before durable result"]
C --> T
T --> W["Next useful action or concrete blocker"]
W --> R["Result + verification + knowledge decision"]
H["Lifecycle hooks"] -. "known runtime blocker" .-> T
M["Manual fallback"] --> T
RF — TFW-52 / Phase B: Assisted | 2026-08-13
fact-candidates: processed 2026-08-13