lidge-jun/dev-pabcd

Structured 5-phase development. Advance only with user approval.

C0/C1 work: see dev §0.0 Work Classifier and §0.1 Patch Fast-Path first — full PABCD is mandatory for C4 and conditional for C3, not the baseline for every task.

dev is canonical: dev §0.2 Rule Classes, §3 Verification Gate, and §5 Safety Rules apply to all work governed by this skill.

§1. Interview Trigger (MUST)

When the user asks for an interview in any form — "인터뷰하자", "인터뷰 모드", "interview", "요구사항 정리", "스펙 정리해줘", "뭘 만들어야 하는지 정리", or any variation — you MUST immediately run:

cli-jaw orchestrate I

Do NOT ask clarifying questions in IDLE instead of entering Interview, skip Interview for unclear requests, or merely narrate — actually execute the command in Bash. /interview is the user-facing shortcut; the Boss agent always runs the command directly.

Interview MUST settle three things before P (DEFAULT, INTERVIEW-CLASSIFY-01): the work class (dev §0.0), the loop archetype (§11.4) — ask "does a verifier define done for this work, or only better?" — and the unit residence (UNIT-RESIDENCE-01, §3.1). Applies in HITL and goal mode alike; an archetype discovered mid-loop, after candidates were burned, is an Interview failure.

§2. How It Works

IDLE ──→ P ──→ A ──→ B ──→ C ──→ D ──→ IDLE
         │      │      │      │      │
        STOP   STOP   STOP   auto   auto
        wait   wait   wait
         └──────┴──── I (Interview) — reachable from any phase, context preserved

Transitions (each phase accepts only its predecessor):
cli-jaw orchestrate I|P|A|B|C|D   → I from any state (context preserved); P from IDLE/I;
                                    A←P, B←A, C←B, D←C (D returns to IDLE)
cli-jaw orchestrate reset         → IDLE from any state (context cleared); re-enter with P

§2.1 Evidence gate (forward transitions)

The four forward transitions (P→A, A→B, B→C, C→D) require an evidence attestation — a real cli-jaw orchestrate command with an --attest JSON, not narration. The server gates the agent (identified by its boss token); a human's /orchestrate X keeps the free pass. P, A, and B also require user approval; C and D proceed automatically once their work is done. In goal mode, §4 rule 4 replaces user approval with evidence-backed checkpoints.

cli-jaw orchestrate A --attest '{"from":"P","to":"A","did":"<the concrete plan you wrote: files/surfaces + devlog path>"}'
cli-jaw orchestrate B --attest '{"from":"A","to":"B","did":"<who audited the plan + the verdict>"}'
cli-jaw orchestrate C --attest '{"from":"B","to":"C","did":"<what you built + who verified it>"}'
cli-jaw orchestrate D --attest '{"from":"C","to":"D","did":"<what you checked>","checkOutput":"<paste the real tsc/test tail>","exitCode":0}'

The gate is form-only: well-formed block, real did narrative (booleans/placeholders rejected); C→D additionally requires non-empty checkOutput and, if present, exitCode:0. It does NOT cross-check the narrative against runtime state — it forces a deliberate, specific claim, not malice-proofing.

Evidence pointers (DEFAULT, ATTEST-EVIDENCE-01): even though the gate checks form only, write did with artifact pointers: plan/devlog path, changed-file list, verifier command, exit code, and relevant cli-jaw goal update checkpoint when goal mode is active, so a later reader can re-check the claim. Narration without running the command does nothing: the state moves only on the command.

Threat model = laziness, not malice. Accepted residuals (NOT bugs): fabricated did, the hidden --force hatch, a prior-turn pendingAttestation, and boss-token stripping (closing it would break the legitimate human-via-CLI free pass).

§3. Phases

P — Plan

If the request has unclear scope or unspecified technology, return to Interview (cli-jaw orchestrate I): present 2–3 options as <TechName> — <plain explanation>, recommend one with project-specific reasoning, confirm once, then proceed.

For broad changes or unfamiliar repositories, P phase MUST include:

• Compact tree of the current repository shape
• Detected repo conventions (docs, plans, architecture notes, naming, tests)
• Whether existing SoT logs (structure/, devlog/, docs/, plans/) were read and will be reused, and whether new ones are proposed
• The SoT sync target (SOT-SYNC-01): which general source-of-truth doc this unit will patch in C — or, if the repo has none, the plan recommends creating one (dev-scaffolding §2.1)

Do not create new project-level source-of-truth folders during B unless approved in P or explicitly requested.

Design phases before mapping them to PABCD. Slice and order phases by dependency/architecture structure (STRICT, PHASE-SPLIT-01) — the orthodox unlimited-time build order: foundations (schema, contracts, core data flow) → core capabilities → integration → hardening/polish — so each phase consumes the verified output of the previous one. DB/API/UI/test work inside a phase are subtasks, and every phase closes with something independently verifiable (build, tests, or a demonstrable surface). Effort-based bucketing is FORBIDDEN (no "quick win vs heavy" buckets, impact/effort matrices, or time-boxed slices): phase boundaries encode build order, not the schedule. Larger work splits into multiple PABCD passes (§5).

Read project docs and dev skills first. Write the complete plan internally, then report it simply — like a developer reporting to the CEO.

Write a plan with two parts: Part 1 — easy explanation of what will be built, in non-developer terms; Part 2 — diff-level precision: exact file paths (NEW/MODIFY/DELETE), before/after diffs for MODIFY, complete content for NEW.

For every planned conditional path, the accept criteria name its activation scenario (C-ACTIVATION-GROUNDING-01 — owner: C phase below).

Loop-spec header (DEFAULT, C2+): open the plan with a compact loop-spec: Loop archetype (§11.4, carried from Interview) · Trigger · Goal (user-visible outcome) · Non-goals · Verifier (the command/gate and what it measures, not only pass/fail) · Stop condition · Memory artifact (worklog/devlog/checkpoint path) · Expected terminal states (§11.2) · Escalation condition. Goal mode also states the §11.5 resource scope.

Instrumentation before candidates: when the archetype is open-ended optimization, the loop-spec MUST include a divergence plan: descriptor axes, cell/archetype assignments, candidate count, deterministic selection rule, and telemetry schema. With a score-only verifier, B's first work item is the telemetry, not a candidate — a scalar-only verifier plus discarded candidates is the §10 plateau signature by design.

If anything is unclear, return to Interview (cli-jaw orchestrate I) — do NOT ask questions in P.

§3.1 Implementation-Unit Documents

Full documentation routine (P concretizes the docs, A audits them as a hard gate, D archives to _fin/, plus the mainstream design-doc/RFC translation table): dev-scaffolding/references/implementation-log.md.

Difflevel roadmap plan (STRICT, DIFFLEVEL-ROADMAP-01): for any multi-phase unit (2+ work-phases), the FIRST P — or the dedicated design-only Phase-0 pass (§5) — must deliver the entire roadmap concretized: 00_plan.md (objective, constraints, dependency-ordered work-phase map) PLUS every phase's decade doc written to full diff-level precision — each a copy-paste-executable PRD, not an outline or empty scaffold. Each later cycle's P re-verifies its pre-written doc against the current codebase (stale check), amends it, then executes. LOOP-CONTINUITY-01 (§10) applies.

Lexicographic separation (STRICT, LEXICO-SPLIT-01): every document in a unit carries a numeric lexicographic prefix — bare semantic filenames (PLAN.md, DIFF_PLAN.md, PHASES.md, RCA.md, unnumbered folders) are an A-phase FAIL, not a style nit. Research/spec material (00-range) and implementation phase designs (decade ranges) are SEPARATE documents; mixing both in one doc fails the audit.

Unit residence (STRICT, UNIT-RESIDENCE-01): every piece of development work belongs to an implementation unit (devlog/_plan/YYMMDD_slug/). Ceremony scales with class (§9); residence does not. C0-C1 fast-path work skips the PABCD ceremony but MUST leave a numbered record doc in its owning unit — next free index in the matching decade — stating what changed, why the fast path applied, and the verification evidence; no owning unit → create a minimal unit folder for it. Interview settles residence before P (§1).

Devlog plan artifacts use decade-range numbering to separate concerns:

| Range | Purpose | |-------|---------| | 00–09 | Research, specs, MOC (00_plan.md, 01_api-survey.md) | | 10–19, 20–29, ... | Phase 1, Phase 2, ... (10_phase1-auth-module.md) |

Rules:

• 00-range durable research is mandatory for C4, and for C3 only when state must persist across turns/agents, public contract or architecture decisions need durable audit, or the user/repo already uses devlog planning; optional for C0-C2 and low-persistence C3 (the work still leaves its numbered record in a unit, UNIT-RESIDENCE-01).
• Sequential within decade; overflow (>10 docs) uses sub-index (00_0_name.md).
• NEVER use bare filenames like PLAN.md, DIFF_PLAN.md, PHASES.md, RCA.md.

Present to the user: the Part 1 summary (≤5 sentences) + diagram + devlog file path, plus "Is there any business logic I must not decide alone?" and "Is this direction correct?"

⛔ Present the plan; revise on feedback. On approval, advance with the canonical P→A attestation form in §2.1.

A — Plan Audit

Spawn a worker to audit the plan (not code). The worker verifies:

• All file paths and imports in the plan actually exist; function signatures match real code; no integration risks; existing source-of-truth docs/logs were read
• No new structure/, devlog/, docs, or AGENTS files are introduced without user approval
• New JS/TS files follow TypeScript preference rules (strict-compatible, or limitations stated) unless the plan states why JS is required
• New devlog phase documents use the numbered lexicographic filename convention; bare-named or research/implementation-mixed docs are a FAIL (LEXICO-SPLIT-01).
• Multi-phase units satisfy DIFFLEVEL-ROADMAP-01: every roadmap phase has a diff-level decade doc (no outline-only or missing phases), and the phase map is dependency-ordered, not effort-bucketed (PHASE-SPLIT-01).
• Each planned conditional path: trigger reachable from states the system actually visits, not consumed upstream first, activation scenario named (C-ACTIVATION-GROUNDING-01). Unreachable-by-construction = plan blocker, not a C-phase discovery.

Output worker JSON for the audit; on FAIL fix the plan and re-audit, on PASS report results to the user.

⛔ Wait for user approval. When approved, advance with the canonical A→B attestation form in §2.1.

B — Build

Implement the plan. You write code by default and own every verdict; a worker may write when its slice passes DISPATCH-ECONOMY-01 (§7.1) and the dispatch is explicitly write-capable with a bounded scope (see Pitfalls). Workers without that grant are read-only verifiers. Do not create structure/ or devlog/ unless approved in P or explicitly requested. After implementing, output worker JSON for verification (code exists, integrates cleanly): NEEDS_FIX → fix and re-verify (repair thresholds §11.3); DONE → report to the user.

⛔ Wait for user approval. When approved, advance with the canonical B→C attestation form in §2.1.

C — Check

Final sanity check: (1) verify all files saved and consistent; (2) run npx tsc --noEmit (TypeScript projects); (3) SoT sync (DEFAULT, SOT-SYNC-01) — patch the repo's general source-of-truth docs (found in P) HERE so SoT and code never diverge silently, or recommend creating one (dev-scaffolding §2.1) in the D summary; (4) report the completion summary.

DEFAULT (C-RENDER-GROUNDING-01): when the work-phase produces an artifact whose correctness only shows when run or rendered (HTML page, SVG, game, UI, chart, animation, script with observable visual/interactive output), C MUST include a render-grounding loop before C→D: (1) RUN it in its natural execution environment — headless-browser screenshot for web, SVG→PNG render, execute scripts, drive games/wizards until the first interactive state change; (2) OBSERVE the output — actually read the screenshot/console back; a produced-but-unread screenshot is not observation; (3) FIX what the observation reveals, then re-run and re-observe. Trigger on artifact type + change, never task depth; stop after ONE clean observation. Static gates (tsc/lint/parse) do not satisfy this rule. Defaults (HEURISTIC): 1280x720 viewport. Evidence scales: C2-C3 record the observation in the attestation narrative; C4 (STRICT) persists the screenshot to the devlog. The render observation is valid checkOutput for C→D and the did references it. Excluded: pure logic/config/prose covered by its own test suite. (Adopted 2026-07-05; devlog 260705_pabcd_render_grounding.)

DEFAULT (C-ACTIVATION-GROUNDING-01): the conditional-path sibling of render grounding. When the work-phase adds/changes a code path that only runs under a trigger absent from the happy path (error handlers, fallbacks, retries, caches, guards, gated branches, mode switches, migration handlers, threshold behaviors), C MUST include activation evidence before C->D: (1) TRIGGER the condition for real (test, scenario, threshold fixture, fault injection); (2) OBSERVE the path execute with its intended effect via an assertion, log line, counter, or trace read back; (3) FIX and re-trigger if observation contradicts intent. "All tests green" does not satisfy this when no test drives the trigger. Retroactive-check signals: byte-identical output to baseline (presume dead, instrument first) and a D claim of "handled/falls back" with no fired-path artifact. The activation observation is valid checkOutput for C->D; the did references it. Excluded: unconditional straight-line changes under existing coverage. P names the activation scenario; A checks trigger reachability.

Long external gates (CI runs, deploys): do not block the turn — register cli-jaw bgtask add --cmd '[...]' --prompt "..." and end the turn; the server re-invokes the boss on completion and PABCD state persists. Local tsc/tests stay blocking.

When done, advance with the canonical C→D attestation form in §2.1; C→D uniquely requires a pasted check tail.

D — Done

Summarize the entire flow: what was planned (P), audited (A), built (B), checked (C); the list of files changed; any follow-up items.

Pessimistic close-out (DEFAULT, LOOP-PESSIMIST-01): for loop/multi-pass work, D also records the negative delta: what did NOT improve, which hypothesis died this cycle, and one sentence answering "what evidence would show the current direction is wrong?" The next P quotes this (§10 LOOP-CONTINUITY-01). D→IDLE→P is a context/bias-flush boundary: the next cycle resumes from disk artifacts, not the transcript's accumulated assumptions.

State returns to IDLE automatically. Project root configuration is persistent: D resets PABCD state but not projectDirs; cli-jaw project clear only on explicit user request.

§4. Rules

1. One phase per response (the gate-and-wait turn boundary; canonical approval rule in §2.1). Goal-mode exception: with an active goal, do not end the turn before D while PABCD-phases remain — keep going P→D, close D, re-enter P for the next work-phase.
2. Sequence: P → A → B → C → D. Use cli-jaw orchestrate reset to restart.
3. Workers verify (read-only) by default; write-capable dispatch follows §7.1 DISPATCH-ECONOMY-01. Verdicts stay with the boss in B.
4. Goal-mode precedence: when a jaw goal is active (dev §0.4), use §2.1 with evidence-backed checkpoints (cli-jaw goal update) instead of user approval; phase order, audit conditions, and verification intensity are unchanged.

Gate quick-reference (strict vs goal mode):

| Gate | Strict PABCD | Goal mode | |------|--------------|-----------| | P→A, A→B, B→C | user approval + --attest | evidence-backed checkpoint + --attest | | C→D | auto + --attest w/ checkOutput/exitCode | same | | Turn boundary | one phase per response | continue P→D within the cycle |

§5. Terminology: work-phase vs PABCD-phase

work-phase = one outcome slice of a larger goal (e.g. "Phase 3: Management API"); PABCD-phase = one letter P/A/B/C/D inside a single orchestration cycle.

Work-phases need not be slices of one feature: successive cycles in the SAME session may target completely different features or plans under the same goal (LOOP-UNIT-CHAIN-01). "This needs its own PABCD" is a plan statement — append the unit to the slice map and run it as the next cycle, never a reason to end the goal or defer to a new session.

Invariant: one work-phase = one full PABCD cycle. Run P→A→B→C→D, close D (→ IDLE), then cli-jaw orchestrate P for the next work-phase. Never run B for several work-phases back-to-back or commit out of B without passing C and D. Depth scales per class (§9); the P→D sequence is never skipped.

Loop / multi-pass tasks: a "loop"/"루프" request (or work too large for one cycle) runs as MULTIPLE PABCD passes — one per work-phase. Pre-plan the full slice map and WRITE all per-phase decade docs (10_phase1, 20_phase2, ...) to diff-level up front (DIFFLEVEL-ROADMAP-01, §3.1; re-verify/amend per cycle). The first pass MAY be a design-only Phase-0 cycle producing exactly this difflevel roadmap. The slice map is APPEND-friendly (LOOP-UNIT-CHAIN-01): an independent unit discovered mid-loop becomes a NEW work-phase appended to the map via a P-phase amendment, then runs as the next cycle in the same session.

Faithful execution (anti-skip): do the real work of each PABCD-phase — P writes the real plan, A really dispatches the audit, B really implements AND verifies, C really runs tsc/tests, D really summarizes with evidence. Never rubber-stamp a phase.

§6. Repository Root Contract

Before writing a PABCD plan or dispatching an employee, determine the actual working repository root with pwd -P from the target repo. If Project root is injected at the top of the system prompt, use it directly; if not, recommend the user configure it (Manager UI → Project settings, or cli-jaw project add /path/to/repo) to avoid JAW_HOME/codebase confusion. Every A/B phase cli-jaw dispatch task body MUST begin with Project root: /absolute/path/to/current/repo.

Rules: Project root is the current working repository, never JAW_HOME; workers never infer the root from ~/.cli-jaw*, process.cwd(), or a temp dir; all relative repo paths resolve against Project root; if it is unknown, STOP and ask before dispatching.

§7. Shared Plan (auto-injected)

When P completes, the plan is saved to the worklog ## Plan section (single source of truth) and kept in ctx.plan; no project-root file is created. In A and B the orchestrator auto-injects the full plan body at the top of every cli-jaw dispatch task under ## Approved Plan, so workers never read a plan file — the task body holds only the audit/verify instruction. Example: write a brief file (Project root: ... + the instruction), then cli-jaw dispatch --agent "Backend" --task-file /tmp/audit-brief.md --async; the result re-enters as employee-results context. Omit --async only for a quick (<2 min) blocking verify. Fan-out: --batch --agents-file <path> --async; overlays: --task-tags "testing,security".

§7.1 Parallel dispatch: isolation, specialists, decorrelated review

• DEFAULT (DISPATCH-ISOLATION-01): Parallel employee lanes in the same work-phase are context-isolated: each dispatch task names an explicit access list — which prior outputs (paths or pasted excerpts) this lane may read — and nothing else from peer lanes is visible. Never paste one lane's in-progress trajectory or draft output into another outside the access list: the first finished trajectory otherwise steers every later lane into redundant agreement (orchestration collapse) and the parallelism buys nothing. This generalizes the §11.8 star topology from divergence lanes to ALL parallel dispatch. Durable cross-cycle artifacts (worklog ## Plan, devlog, D summaries) stay shared; widening an access list is a stated decision in the dispatch task.
• HEURISTIC (SPECIALIST-CRUX-01): When P or A identifies a narrow crux outside the main line's domain — a derivation, a protocol subtlety, a domain constant, a security property — dispatch a specialist lane to re-derive that crux from first principles before merging the work that depends on it. The specialist's return names its assumptions, the derivation or trace, and the exact decision its verdict changes.
• HEURISTIC (REVIEW-DECORRELATE-01): When the runtime offers more than one model family, run the A reviewer — and any §11.3 clean-slate re-examination after repeated failed repairs — on a different model family than the one that produced the plan or build. Same-family reviewers share blind spots; decorrelating the reviewer is the cheapest independence upgrade available.
• DEFAULT (DISPATCH-ECONOMY-01): Decide delegation on three axes — specifiability (can the task packet state its inputs, outputs, and decision boundary completely?), verifiability (can the return be checked mechanically: anchors, figures, reproduction commands?), and judgment ownership (collapse/crux VERDICTS stay with the boss; re-derivation, standardized implementation, research, and audits are dispatchable). A slice that fails any axis stays with the boss. Complexity or importance is NOT an axis: a hard derivation that passes all three axes is dispatchable (SPECIALIST-CRUX-01 above), and a trivial slice whose packet cannot state its decision boundary is not. Output side: every return receives an accept/reject/merge disposition with a one-line rationale BEFORE the next wave is dispatched (wave-granular allowed). Returns must carry verbatim anchors — file:line quotes, exact figures, URLs — a paraphrase-only summary creates correlated blind spots between boss and reviewer. Prefer batch-waves with a single synthesis over drip-feed spawning. Speculative dispatch of a later phase's work is default-OFF; the one exception is phase-invariant EXTERNAL research (reads no repo state), quarantined as candidate — unverified and discarded by default on plan amendment.

(Adopted 2026-07-07 from the Sakana Fugu learned-orchestrator report, arXiv:2606.21228; canonical record: pabcd_initiative devlog 260707_fugu_orchestration_adoption. DISPATCH-ECONOMY-01 adopted 2026-07-11 from an adversarial fork-debate + Tier-2 arXiv claim ledger (10 papers, evidence grades recorded; codexclaw devlog 260711_dispatch_economy_docs_site).)

§8. Pitfalls

Delegation Trap — B phase: Boss writes by default. Workers are READ-ONLY verifiers unless granted otherwise. A worker may write only when the slice passes DISPATCH-ECONOMY-01 (§7.1: specifiable, verifiable, no verdict ownership) AND the dispatch is explicitly write-capable with a bounded scope (--mutable, optionally --scope). Without --mutable, "implement/write/create" tasks are forbidden. Always allowed: "verify src/x.ts compiles", "check integration of Y", "report DONE or NEEDS_FIX".

Context Drift — a worker saying "I'll proceed based on my assumption of the plan" → STOP; verify the dispatch went through /api/orchestrate/dispatch (only that path auto-injects the plan). Never let workers reconstruct the plan from a short task.

Phase Skip — A (audit) is mandatory for C4, and for C3 when public contract, architecture, persistence, cross-agent, or cross-session risk exists; micro-audit for C2, optional for C0-C1 (dev §0.0). B verification is never "skippable"; intensity scales with class (PABCD-AUTO-01). The orchestrator does not enforce these gates — YOU do.

§9. PABCD Depth by Work Class

| Class | Plan (P) | Audit (A) | Build (B) | Check (C) | Record (D) | |-------|----------|-----------|-----------|-----------|------------| | C0-C1 | None/inline | Optional | Direct fix | Smallest proof | One-line summary as a numbered record doc in the owning unit (UNIT-RESIDENCE-01) | | C2 | Compact plan | Micro-audit | Boss-led build, focused tests | Targeted gate | Summary | | C3 | Compact or full PABCD plan depending on persistence/risk | Required when public contract, architecture, persistence, cross-agent, or cross-session risk exists; otherwise focused audit | Boss-led build (economy-eligible slices dispatchable, §7.1), employees verify only when useful | Affected suite + docs consistency when docs/contracts changed | Summary + evidence; durable record only when state must persist | | C4 | Full PABCD plan (mandatory) | Required, independent | Boss-led build, employee verifies | Full relevant gates | Durable risk/approval/evidence record | | C5 | Interview/research first | — | — | — | Reclassify, then follow the new class |

Render-artifact work-phases add C-RENDER-GROUNDING-01 (§3 C) to the Check column at C2+; C4 escalates its evidence to STRICT (persisted screenshot).

§10. Optimization-Loop Meta-Rules (plateau discipline)

These rules apply to score/objective-maximization loops and repeated PABCD passes where candidates are being discarded by evidence gates. Gate validity itself is owned by dev-testing §9.5 Limited-Oracle / Score-Objective Evaluation.

• DEFAULT (LOOP-PHASE-DEATH-01): Track each discarded candidate's killing PABCD-phase and change class (parameter-tweak, branch-toggle, state-space redesign, evaluator change). After N consecutive same-phase, same-class deaths (start N=3 — HEURISTIC, tune per domain), the next work-phase MUST target the killing mechanism itself — usually the evaluation gate — not another candidate of that class.
• STRICT (LOOP-CONTINUITY-01): P must begin by quoting the previous cycle's D conclusions and next-direction. A new candidate that contradicts the recorded next-direction requires an explicit stated reason.
• DEFAULT (LOOP-CANDIDATE-ANCHOR-01): For score/objective-maximization work, source divergence candidates from domain-state evidence such as logs, trajectories, and opponent/instance analysis, not only from existing code parameters. All-lever-tweak candidate sets are parameter-space anchoring: regenerate from the state space.
• HEURISTIC (LOOP-INSTANCE-CHECK-01): Check whether evaluation instances are fixed and enumerable: fixed opponents, fixed test maps, fixed graders. If yes, per-instance specialization (fingerprint + playbook) is a legitimate widening move; consider it before generic-strategy tweaks.
• DEFAULT (LOOP-MECHANISM-PROOF-01): A candidate whose value is a new branch or mechanism must carry activation evidence from the instances it targets: a counter, debug line, or trace showing the branch actually fired WITH its intended effect before adoption. Aggregate score movement is not activation proof; in a multi-feature combo a dead mechanism hides behind other features' gains, so each branch needs its own trace. Loud special cases: a zero-delta ablation means presume it never ran and instrument before combining or discarding; byte-identical outcomes signal a dead path, while a branch that runs and loses should still move some trace detail.
• DEFAULT (LOOP-RESIDUAL-TRACE-01): A residual failure carried through D needs a mechanism-level explanation: which branches fired, which did not, and why the outcome followed; otherwise label it unexplained. A plausible opponent/environment story is not evidence unless the trace confirms our own mechanism armed and acted.
• HEURISTIC (LOOP-PEER-CONTRAST-01): When a peer's bot, reference solution, or competitor run achieves the objective on a fixed instance we fail, the next generation's first analysis deliverable is the behavioral diff of the two traces before any new candidate — the cheapest capability-gap detector.
• HEURISTIC (LOOP-FANOUT-TIMING-01): Spend parallel fan-out late, not early. While coarse levers still move the metric, stay single-track (N=1); once coarse levers stabilize or plateau and the search shifts to fine-grained candidates, parallel candidate lanes and specialist re-derivation start paying for their cost. Fan-out also buys outcome consistency (cross-run variance reduction), not only peak score. (Adopted 2026-07-07 from Sakana Fugu, arXiv:2606.21228: orchestration gains on a 123-experiment autonomous training loop concentrated after mid-run, once coarse configuration search gave way to fine optimizer/schedule tuning.)

Grounding: a 14-discard plateau where a prefix-only replay gate and a hard invariant locked a 3.5/8 score. Single-incident induction: treat constants as starting values and revise when a second domain contradicts them.

§11. Loop-Engineering Alignment

PABCD is the macro loop; loop engineering supplies the inner-loop rules for prompts, verifiers, repair, exploration, and resource bounds inside each phase.

§11.1 Loop values (DEFAULT)

• Feedback must change the next action. A result that does not alter the next step is a retry, not a loop; read the failure delta first.
• The verifier outranks the prompt. Prefer deterministic evidence (tests, exit codes, diffs, telemetry) over model self-assessment; use cli-jaw dispatch employees as independent verifiers when the class/risk warrants it.
• Memory lives on disk, not only in the transcript: worklog ## Plan, devlog, attestations, goal checkpoints, death logs — the next iteration resumes from artifacts.
• Budget exhaustion is not done. Never report a budget stop as success.
• Context pressure is not budget exhaustion. Compaction is survivable BY DESIGN because memory lives on disk: an approaching context limit means checkpoint durable state (worklog, goalplan, devlog) and continue after the flush — never grounds to close the goal, shrink the plan, or report DONE/BUDGET_EXHAUSTED. BUDGET_EXHAUSTED requires a bound the plan actually stated (tokens, cost, wall-clock).
• Interview does not solve intent transfer. It yields an initial loop-spec; later evidence may force cli-jaw orchestrate I — intent refines cyclically.

§11.2 Terminal-state vocabulary (DEFAULT)

D is the success exit, not the only exit. D summaries must name the actual report state: DONE (verified success), NOOP (nothing needed), BLOCKED (external dependency), UNSAFE (human risk decision), NEEDS_HUMAN (user-only judgment), or BUDGET_EXHAUSTED (adopt best-so-far and say so). Report states, not extra FSM states: cli-jaw still closes via D or reset.

§11.3 Repair-loop discipline (C→B returns)

The B/C inner loop is: implement → run verifier → read the failure delta → repair only the failing delta → re-verify. DEFAULT (LOOP-REPAIR-01): 2 consecutive failed repairs of the same failure → stop patching, enter root-cause mode (dev-debugging); 3 → replan (cli-jaw orchestrate P) or return to Interview. HEURISTIC (LOOP-DOOM-01): 3 attestation failures in the same phase within one work-phase means no progress; force an Interview return — the server may not enforce this, so the Boss agent self-applies it.

§11.4 Loop archetype by problem type (DEFAULT, LOOP-ARCHETYPE-01)

Classify the work-phase's problem type before choosing the inner loop shape: Spec-satisfaction (verifier defines done — tests, contracts, npx tsc --noEmit): use the §11.3 repair loop; it can converge. Open-ended optimization (verifier defines only better — scores, win rates, adversarial opponents): repair loops plateau; use an explore-and-select loop — generate diverse candidates, evaluate on the same instances, keep best-so-far, regenerate from the winner, stop on plateau (§10 LOOP-PHASE-DEATH-01) or budget, terminal state BUDGET_EXHAUSTED with best adopted, not DONE. A repair loop on an optimization problem is a category error; fix the loop shape, not the cycle count.

§11.4a Analysis-before-regeneration (DEFAULT, LOOP-REANALYZE-01)

Repair fixes actions; analysis revises the model that generates them. In an explore-and-select loop, every generation MUST begin with an analysis deliverable: (1) an updated problem/opponent model from evidence (telemetry, replays, failure deltas) and (2) capability-gap hypotheses — what the artifact cannot yet sense or do; a gap hypothesis may expand the allowed patch surface, which is a P-level amendment. Candidates are sourced from these hypotheses (§10 LOOP-CANDIDATE-ANCHOR-01) and the next P quotes them (§10 LOOP-CONTINUITY-01). Regenerating straight from scores is a repair loop wearing an explore-and-select label.

§11.4b Mechanism activation proof (DEFAULT, LOOP-MECHANISM-PROOF-01)

Scores select candidates; only traces verify mechanisms. Owner: §10 LOOP-MECHANISM-PROOF-01 (activation evidence with intended effect, combo masking, zero-delta/byte-identical semantics). Grounding (NEXT NATION 2026-07): a structurally unreachable branch shipped inside a passing combo, recorded as "weak" instead of instrumented; one trace line at adoption time would have caught it a work-phase earlier.

§11.5 Unattended-loop resource policy (DEFAULT)

Goal-mode loop-specs must state tool/credential scope, token/cost budget, and wall-clock bound; record resource decisions in the plan and cli-jaw goal update checkpoints. For C4 surfaces an unattended loop with unstated scope is an ESCALATE-class omission: stop and ask before running it.

§11.6 Continuation doctrine (DEFAULT, LOOP-CONTINUE-01)

The loop keeps the turn alive; the agent decides what "remaining work" means. On loop re-entry or after a D close: do not redefine the objective downward (P/goalplan criteria are the bar); audit completion against current repo state, not memory; read durable state first (worklog/devlog/goalplan ledger) to recover remaining work-phases/criteria; and IDLE is not the end while work remains — under an active goal, start the next work-phase at P. Work-phases chain HETEROGENEOUS units in one session (LOOP-UNIT-CHAIN-01): an independent feature discovered mid-loop is appended to the plan and started at P, not deferred to a new session or used to justify closing the goal.

§11.7 Divergence/collapse (DEFAULT)

PABCD is convergence-first by default. For ordinary build or bug-fix goals, keep one strategy. Divergence is a mode for the open-ended-optimization archetype (§11.4), not a standing habit. Entry: deliberate in HITL (during I/P when intent is open, the approach is genuinely uncertain, the objective is maximize/deceptive, or the user asks to compare); in goal mode, prompted on plateau detection (non-improving metrics).

When divergence is ON: record at least two candidates with search provenance; collapse early at P for spec-satisfaction work (pass/fail, locally checkable); collapse late at D for maximize-metric work where the local metric can deceive (build candidates in isolation, evaluate on the same instances, keep/discard by recorded metric); after the plateau breaks or a candidate is kept/discarded, turn divergence off (N=1 loop). Divergence never bypasses human confirmation in HITL; in goal mode the hook only keeps the turn alive and tells the agent to re-plan — never asking or moving phases.

§11.8 Divergence cost tiers (DEFAULT, DIVERGE-TIER-01)

Divergence (§11.7) defaults to CONCEPTUAL candidates, not implemented ones. Choose the cheapest tier that can kill the wrong option:

• Tier 0 — inline brainstorm. The Boss session itself lists options with trade-offs inside the plan/interview. No dispatch. Default for ordinary uncertainty.
• Tier 1 — conceptual candidate docs (the divergence default). 2-3 parallel cli-jaw dispatch employee lanes each produce ONE one-page candidate direction doc (no code, no worktrees) with mandatory front-matter: assumptions, risks, kill-criteria, evidence-needed. Lane research stays read-only; the doc write is scoped to the worklog/devlog plan archive. The BOSS session (collapse owner) performs critique/triage directly — it holds the most context; a separate cross-critique round is waste. Collapse gate: N candidate docs with filled front-matter AND per-candidate provenance — search provenance per §11.7 for externally-sourced candidates; for candidates grounded in the local codebase a repo-evidence path is acceptable, an EXPLICIT §11.8 AMENDMENT to §11.7's search-only wording. The mandatory front-matter keeps the gate otherwise stricter than §11.7. Cross-critique rounds are NOT a gate condition.
• Tier 2 — implementation spike (rare escalation). Parallel worktree implementations judged by the same verifier, ONLY when both hold: (a) the choice is load-bearing and Tier-1 candidates genuinely conflict on it, and (b) judging requires running code (performance assumptions, live API contracts, deceptive local metrics). Expected frequency: 0-1 per unit. Tier-2 entry is a recorded P-level decision in the worklog/devlog plan.

Budget rationale: employee tokens may be near-free, but wall-clock and the Boss's triage attention are not. Tier inflation (defaulting to Tier 2 because employees are cheap) is a discipline violation; so is tier deflation that lets a load-bearing conflict collapse from paper arguments alone.

Unknowns lane: the first Tier-1 dispatch of a research-heavy or unfamiliar-surface unit SHOULD be a blindspot/unknowns pass (known unknowns, unknown knowns recoverable from references, unknown unknowns from codebase/web search), so candidates are sourced from evidence, not parameter tweaks (§10 LOOP-CANDIDATE-ANCHOR-01).

Topology: candidates and critiques never flow employee-to-employee; exchange is file-mediated through the worklog/devlog archive, and the Boss schedules rounds and owns the collapse (star-shaped exchange). Employees MAY use their own CLI sub-agents internally per the agent-spawn doctrine; that internal parallelism does not change the star-shaped candidate exchange or move collapse ownership.

§11.9 Crux-matched collapse (DEFAULT, COLLAPSE-AGGREGATOR-01)

Collapse and synthesis ownership follows the disputed crux. When parallel candidates disagree, the synthesis verdict comes from whoever is strongest on the domain of the disagreement — the Boss dispatches a crux-matched aggregator lane when it is not itself strongest there (the lane returns a verdict; the Boss still owns the collapse decision per §11.8 topology). A fixed aggregator caps the exercise at that aggregator's own ceiling for the domain. The collapse record names each candidate's partial correctness, each candidate's failure mode, and why the winner's evidence resolves the crux. (Adopted 2026-07-07 from Sakana Fugu, arXiv:2606.21228.)