00_plan — Implementation-unit discipline as the harness default (de-branded, strict)

Source: devlog/_fin/260705_difflevel_roadmap_default/00_plan.md

00_plan — Implementation-unit discipline as the harness default (de-branded, strict)

Date: 2026-07-05 (rev 3 — scope expanded to 4-harness roadmap by user directive)

Unit: devlog/_plan/260705_difflevel_roadmap_default/

Class: C3 (cross-agent contract change to the canonical methodology spec; durable audit trail required)

Loop-spec

• Loop archetype: spec-satisfaction (verifier defines done)
• Trigger: user interview 2026-07-05 — the harness spec does not enforce the

difflevel-plan routine; agents keep producing research-only units, effort-bucketed

phase splits, bare-named plan files, and undocumented small work. Root causes: the

routine is branded ("Jawdev style") and therefore opt-in; deferred concretization

("fill per cycle") is the default; no effort-bucketing ban; unit residence gated on

work class.

• Goal: any agent following the pabcd_initiative skill family must, without being

asked: (1) treat implementation-unit documentation as the unbranded DEFAULT process,

(2) for multi-phase work produce the full roadmap with EVERY phase doc at diff-level

up front, dependency-ordered, (3) never write bare-named or mixed research/implementation

docs, (4) leave a numbered record in a unit for ALL work, down to C0 hotfixes.

• Non-goals: FSM/server code changes; renumbering existing unit folders;

blind-copying text across harnesses (each port adapts to that repo's conventions).

• Verifier:
• grep -ric jawdev skills/ → 0 hits (full de-brand)
• DIFFLEVEL-ROADMAP-01 / PHASE-SPLIT-01 / LEXICO-SPLIT-01 / UNIT-RESIDENCE-01

present at definition site (dev-pabcd §3.1/§3) and at every consumer touchpoint

listed in 10_phase1

• grep -rn "fill per cycle" skills/ → 0; grep -n "scaffold per-phase decade docs" skills/dev-pabcd/SKILL.md → 0
• independent A-phase audit (opus-4-6 per user instruction) + B read-back
• Stop condition: all MODIFY blocks in 10_phase1_harness_patch.md applied and

verified; no files outside the 4 listed touched.

• Memory artifact: this unit folder.
• Expected terminal states: DONE, or NEEDS_HUMAN if a gate decision is rejected.
• Escalation condition: any patch target text no longer matches (files drifted).

Evidence baseline (why the harness fails today)

1. Branding = opt-in escape hatch: dev-scaffolding/SKILL.md gates the routine on

"The user asks for Jawdev/Lidge-style structure" (3 places). A named style invites

"the user didn't ask for that style" as a skip reason.

1. Deferred concretization is the default: implementation-log.md L37-38 "written

in P of ITS cycle (pre-scaffold … fill per cycle)"; dev-pabcd §5 "scaffold

per-phase decade docs up front" — empty scaffolds satisfy the letter of the rule.

1. No effort-bucketing ban: 0 grep hits across the skill family.
2. Unit residence gated on class: implementation-log.md "C0-C2 work does not get

a unit folder"; dev §0.1 keeps change documentation only "when a worklog/changelog

file is provided".

1. Field evidence (opus-4-6 scan of 12 recent units, 2026-07-05): newest unit

codexclaw/…/260705_hook_diet_skill_implicit = master plan + 6 research docs,

0 phase docs, 0 diff markers; bare-name violations still occur

(chatcli/devlog/_plan/mvpplan/, historical PLAN.md/DIFF_PLAN.md).

Good exemplar for the target state: 260703_gui_production_hardening

(8 phase docs, 27 NEW/MODIFY/DELETE blocks).

1. Downstream reference survey (for later port units, out of scope here):

cli-jaw skills_ref has identical weaknesses incl. the 3 opt-in gates; codexclaw

partial (decade numbering present, no upfront/ban rules); jawcode jwc family has

no implementation-unit concept at all (6 skills, no devlog/numbering rules).

Interview decisions (2026-07-05, binding)

1. Depth: ALL phases fully diff-level up front; later cycles' P does a stale

re-verify of the pre-written doc, never writes it fresh.

1. Split criterion: dependency/architecture order (foundations → core →

integration → hardening). Effort bucketing FORBIDDEN.

1. Nature of fix: harness spec improvement — the spec text enforces this, not

per-request triggers.

1. Scope (rev 3, supersedes rev 1-2): full 4-harness roadmap in THIS unit —

Phase 1 pabcd_initiative (canonical), then cli-jaw, jawcode, codexclaw, each

adapted to its own conventions, INCLUDING system-prompt and hook/config surfaces

that inject planning/devlog rules, not only skill files.

1. De-brand: remove the "Jawdev" name entirely; the routine becomes anonymous

STRICT rules — a named style is skippable, a rule is not. Interview phase also

settles unit residence, not just class/archetype.

1. Lexicographic strictness: bare plan filenames are an A-phase FAIL; research

(00-range) and implementation phase designs (decade ranges) are strictly separate

documents.

1. Universal unit residence: every piece of work lands in an implementation unit.

Ceremony scales down (C0/C1 skip PABCD), residence never does — a hotfix leaves a

numbered record doc (e.g. 40_hotfix_slug.md) in its owning unit stating what was

done and why the fast path applied.

1. SoT sync (rev 4): before patching a repo, FIND its general source-of-truth

docs (structure/, architecture/INDEX) first; units that change

architecture/contracts patch the SoT doc in the SAME unit (C gate); if the repo

has no SoT doc, RECOMMEND creating one.

New rule IDs introduced

ID	Class	One-line summary	Definition site
DIFFLEVEL-ROADMAP-01	STRICT	Multi-phase unit ⇒ first P writes EVERY phase doc to diff-level (copy-paste PRD); later P = stale re-verify + amend	dev-pabcd §3.1
PHASE-SPLIT-01	STRICT	Phases sliced/ordered by dependency/architecture; effort bucketing (quick-win/heavy, impact-effort) forbidden	dev-pabcd §3 P
LEXICO-SPLIT-01	STRICT	Every unit doc numbered (lexicographic); bare names FAIL the audit; research and implementation designs never mix in one doc	dev-pabcd §3.1
UNIT-RESIDENCE-01	STRICT	All work belongs to an implementation unit; C0-C1 skip ceremony but MUST leave a numbered record doc	dev-pabcd §3.1
SOT-SYNC-01	DEFAULT	Find the repo's general SoT docs first; patch them in the same unit at C; recommend creating one if absent	dev-pabcd §3 C / dev-scaffolding §2.1

Work-phase map (dependency-ordered)

Order rationale: canonical rule text must land first (Phase 1 defines the rule IDs

and wording every port adapts); cli-jaw is the live-deploy source the user works in

daily (Phase 2); codexclaw already partially adopts the routine so its port is a

delta (Phase 3); jawcode lacks the concept entirely and needs the largest adapted

surface, informed by all prior ports (Phase 4). One full PABCD cycle per phase.

Phase	Doc	Outcome	Verifier
1	10_phase1_harness_patch.md	pabcd_initiative: 4 files patched — rule definitions (dev-pabcd), routine spec (implementation-log), scaffolding consumer (dev-scaffolding), router fast-path (dev); de-branded and hardened	grep assertions above + read-back audit
2	20_phase2_cli_jaw.md	cli-jaw: same contract ported into skills_ref + any system-prompt/hook surface that injects planning rules, per cli-jaw conventions	per-doc grep assertions + read-back audit
3	30_phase3_codexclaw.md	codexclaw: cxc/pabcd skill family + plugin hooks/prompt surfaces hardened (delta port — partial adoption exists)	per-doc grep assertions + read-back audit
4	40_phase4_jawcode.md	jawcode: jwc skill family + role agent prompts/defaults gain the implementation-unit contract (largest adaptation)	per-doc grep assertions + read-back audit

Docs 20/30/40 are written to diff-level in THIS first P (DIFFLEVEL-ROADMAP-01),

from per-repo touchpoint inventories (opus-4-6 agents, exact paths + verbatim

quotes). A audits in two rounds: round 1 = 00+10 (canonical), round 2 = 20/30/40

once the inventories land. Later cycles' P re-verifies each pre-written doc for

staleness before B, per the rule itself.

Risks

• Semantic reversal (carried from rev 1): PHASE-SPLIT-01 replaces the

"user-visible outcome unit" default and deletes "Do not plan a whole database/API

foundation as one PABCD before any usable outcome exists." Intentional (user: 정석 /

unlimited-time process); mitigated by "every phase closes verifiable".

• Doc staleness: all-upfront diffs go stale as earlier phases land. Accepted;

mitigated by mandatory stale re-verify in each later cycle's P.

• Record-doc overhead: UNIT-RESIDENCE-01 adds one small numbered doc per hotfix.

Accepted by user (rev-2 interview); the rule text keeps the record minimal

(what/why-fast-path/evidence — not ceremony docs).

• Cross-repo reference drift: downstream harnesses still reference "§3.1 Jawdev

Document Numbering" by its old heading; those repos are out of scope and will be

fixed in their own port units. Within THIS repo all internal references are updated

in the same patch.

Attestation log

• 2026-07-05 P (rev 1): plan written; FSM server down ("port undefined") — per the

agent-neutral runtime adapter, transition announced in reply and attested here.

{"from":"I","to":"P","did":"interview settled 4 binding decisions; wrote 00_plan.md + 10_phase1_harness_patch.md with 8 MODIFY blocks at diff-level"}

• 2026-07-05 P (rev 2): user feedback expanded scope within the repo (de-brand,

LEXICO-SPLIT-01, UNIT-RESIDENCE-01); downstream survey attached as evidence; both

plan docs rewritten.

• 2026-07-05 P (rev 3) + P→A: user approved and expanded scope to the 4-harness

roadmap ("pabcd 돌고 첫번째는 initiative에만, 그다음 cli-jaw/jawcode/codexclaw 전부,

시스템 프롬프트·hook 설정 포함"). FSM server still down — attested here per the

runtime adapter.

{"from":"P","to":"A","did":"rev-3 plan: 00_plan.md 4-phase dependency-ordered map + 10_phase1_harness_patch.md (4 files, 21 diff-level blocks); dispatched opus-4-6 audit (round 1: 00+10) and 3 parallel opus-4-6 port-touchpoint inventories for phases 2-4 docs"}

• 2026-07-05 A round 1 (phase 1): opus-4-6 audit verdict PASS, 0 blockers, 3 warns

+ 1 note. All warns incorporated into the plan as blocks 1.6 (§9 depth-table record

cell), 3.10 (folder-approval gate exemption), and a fixed verification grep. Note

(self-compliance: A entered before 20/30/40 existed) acknowledged — round 2 audits

those docs before any phase 2-4 B.

{"from":"A","to":"B","did":"opus-4-6 worker audited 00_plan.md + 10_phase1_harness_patch.md against repo: PASS with 3 warns, all folded into plan rev 3 (blocks 1.6/3.10 + grep fix)"}

• 2026-07-05 B (phase 1): all 23 blocks applied to the 4 target files

(skills/dev-pabcd/SKILL.md, skills/dev-scaffolding/references/implementation-log.md,

skills/dev-scaffolding/SKILL.md, skills/dev/SKILL.md). Doc deviations (audit warns)

were edited into 10_phase1 BEFORE coding past them, per the routine.

{"from":"B","to":"C","did":"applied 23 MODIFY blocks across 4 skill files in pabcd_initiative; plan doc amended first (blocks 1.6/3.10)"}

• 2026-07-05 C (phase 1): verification greps all green —

`jawdev hits: 0 files · DIFFLEVEL dev-pabcd:3 · LEXICO dev-pabcd:2 · UNIT-RES dev-pabcd:4 ·

PHASE-SPLIT dev-pabcd:2 · UNIT-RES impl-log:2 · UNIT-RES dev-scaffolding:2 · UNIT-RES dev:1 ·

fill-per-cycle:0 · pre-scaffold:0 · scaffold-EOL dev-pabcd:0`

{"from":"C","to":"D","did":"ran the 10_phase1 verification grep suite","checkOutput":"jawdev:0 files; DIFFLEVEL:3; LEXICO:2; UNIT-RES:4/2/2/1; PHASE-SPLIT:2; fill-per-cycle:0; pre-scaffold:0; scaffold-EOL:0","exitCode":0}

• 2026-07-05 D (phase 1): DONE. Planned (rev 1-3 interview-driven), audited

(round-1 PASS + warns), built (23 blocks), checked (grep suite green). Changed:

4 skill files. Negative delta / pessimistic close-out: the FSM server being down

means no server-side gate exercised these attestations — they are worklog-only;

evidence that the direction is wrong would be a future unit where all-upfront

diff docs went stale faster than the stale-recheck could absorb. Next: phases 2-4

port docs (20/30/40) authored + audited, then one PABCD cycle per port.

• 2026-07-05 P (port docs) + A round 2: three opus-4-6 authors wrote 20/30/40 from

the patched canonical + direct repo inspection; three independent auditors ran per

doc. Verdicts: cli-jaw FAIL (3 blockers: file-count header, JDS-003 rename

missing from test rewrite, verification-count self-inconsistency), codexclaw

FAIL (1 blocker: "fill per cycle" verification self-hit — the prohibition

sentence trips its own anti-regression grep), jawcode PASS (2 warns: dropped

PHASE-SPLIT subtask/closure clause, contradictory snapshot-test verification

line). Cross-cutting lesson: two of three ports independently dropped the same

canonical subtask/closure clause — flagged for future rule-sync tooling.

FAIL → fix → re-audit loop dispatched (fix all three docs; re-audit the two

FAILs), per the A-phase repair rule.

• 2026-07-05 A round 3: fix agents resolved all findings (cli-jaw now 13 files incl.

structure/ de-brand + HTML-regen deploy note; codexclaw verification self-hit fixed,

PHASE-SPLIT clause restored, style normalized; jawcode warns applied). Re-audits:

cli-jaw PASS (0 findings), codexclaw PASS (0 findings), jawcode PASS carried.

All three port docs are now approved P artifacts.

• 2026-07-05 P amendment (rev 4, user directive): SOT-SYNC-01 added — find general

SoT docs first, patch them in the same unit at C, recommend creation when absent.

Blocks 1.7/1.8/2.5/3.11 added to 10_phase1 and applied to canonical (B delta);

C greps: SOT dev-pabcd 2, dev-scaffolding 1, impl-log 1, jawdev still 0.

Port docs receive equivalent SOT-SYNC-01 blocks before their B execution.

• 2026-07-05 B (phases 2-4, opus-4-6 executors + boss trailing observation):
• cli-jaw: 13 files (skills_ref ×5 incl. goal, src ×5, test rename

jawdev→devlog-skill-contract with JDS→DLC ×3, structure ×2). One legit B repair:

builder.ts guide line shortened to fit test PSC-004's 1600-char window (both

rule IDs kept); plan doc block 6.1 synced to the applied text afterward.

Gates: tsc exit 0; gate:all 12/12; renamed test 3/3; npm test failures = 2

pre-existing in untouched files. dev-pabcd skills_ref blocks were pre-applied

by the parallel render-grounding session — verified verbatim, no conflict.

• codexclaw: only dev/SKILL.md ×3 needed application (pabcd + dev-scaffolding

+ implementation-log pre-applied identically by the parallel session). Build 84

files exit 0; tests 687/687.

• jawcode: 4 prompt/skill files, zero drift, all 6 blocks verbatim; target

tests pass (jaw-interview-skill-policy, workflow-surface-orchestrate, snapshots

29/29); check:ts/test:ts failures all pre-existing in unrelated TS sources.

{"from":"B","to":"C","did":"three opus-4-6 executors applied 20/30/40 to their repos; boss trail-reviewed every runtime-prompt diff live (state-machine.ts, builder.ts, planaudit.ts, orchestrate-i/p/c.md, structure docs)"}

• 2026-07-05 C (phases 2-4) + polish (boss, direct):
1. jawcode orchestrate-c.md header said "four-stage" while the body numbers

Stage 1-5 — polished to "five-stage … SoT sync → render grounding

(conditional) → verdict"; 40_phase4 block 4.1 + greps synced.

1. 40_phase4 verification self-hits corrected: "fill per cycle"/"quick win"

expect 1 (the prohibition sentences themselves), not 0.

1. 20_phase2 block 6.1 After synced to the applied PSC-004-window text

(doc-code convergence).

1. SOT-SYNC-01 dogfood: pabcd_initiative README.md (this repo's SoT) patched

with the 2026-07-05 hardening record. cli-jaw structure/ docs were patched in

phase 2; codexclaw has no structure/ SoT (recommendation: create one — noted);

jawcode SoT surfaces are the prompt files themselves.

Final cross-repo greps: jawdev = 0 on all target surfaces (canonical, cli-jaw

skills_ref+src+tests+structure, codexclaw skills, jawcode jwc+prompts — the two

jawcode "hits" are the cliJawDevSkills camelCase template variable, a false

positive, not the brand); SOT-SYNC-01 present at all five runtime surfaces;

jawcode header five-stage 1 / three|four-stage 0.

{"from":"C","to":"D","did":"final cross-repo grep suite + per-repo gates","checkOutput":"jawdev 0/0/0/0 (2 camelCase false positives); SOT-SYNC-01 2/2/1/2/1; five-stage 1, three|four-stage 0; tsc 0, gate:all 12/12, codexclaw 687 pass, jawcode target tests pass","exitCode":0}

• 2026-07-05 D (unit close-out): DONE — all four phases executed and verified.

Changed: canonical 5 files (4 skills + README SoT), cli-jaw 13, codexclaw 4 (3 by

parallel session, 1 file this unit's executor), jawcode 5 (4 planned + header

polish). NOT done (user-owned follow-ups): commits in all four repos (working

trees left dirty for review, per no-proactive-git rule); cli-jaw deploy

(npm run build + jaw skill sync + server restart) and docs/dev HTML

regeneration; codexclaw ships on next session load; jawcode on next build.

Negative delta (pessimistic close-out): FSM server was down the whole unit, so no

server-side gate exercised any attestation — worklog-only discipline; the

parallel render-grounding session pre-applying blocks in two repos means our

"pre-applied" verification trusted verbatim-match, not independent application;

the cliJawDevSkills variable keeps a case-insensitive jawdev grep from ever

reaching literal 0 in jawcode. Evidence that would falsify the direction: a

future multi-phase unit where upfront diff docs went stale faster than the

stale-recheck absorbed, or record-doc fatigue on C0-C1 hotfixes.

Other files in this unit

• 10_phase1 — Harness patch: de-brand + DIFFLEVEL / PHASE-SPLIT / LEXICO / UNIT-RESIDENCE (10_phase1_harness_patch.md)
• 20_phase2 — cli-jaw port: de-brand + DIFFLEVEL / PHASE-SPLIT / LEXICO / UNIT-RESIDENCE / SOT-SYNC (20_phase2_cli_jaw.md)
• 30_phase3 — Codexclaw port: DIFFLEVEL / PHASE-SPLIT / LEXICO / UNIT-RESIDENCE / SOT-SYNC + de-brand (30_phase3_codexclaw.md)
• 40_phase4 — Jawcode port: implementation-unit discipline into the jwc skill family (40_phase4_jawcode.md)