harness: codify the seed run — discovery-to-orchestration v3 run shape

[rickylabs]

What

Codifies the agentic pipeline that produced plan-roadmap-expansion--seed (PR #397; board
#399–#461) as a reusable Harness v3 run shape: the seed run.

• .llm/harness/workflow/seed-run.md — the profile. Stage contracts A–I (bootstrap → discovery
  corpus → synthesis → deep-dive packs → plan lock → adversarial → PLAN-EVAL → owner ratify +
  one-shot filing → handoff), lane bindings by reference to lane-policy.md only, evidence-citation
  gate, drafts-only-until-ratification boundary, GitHub-authority-after-filing rule, scale-to-fit +
  "when NOT to use", landmine pointers, dogfood acceptance criterion.
• .llm/harness/templates/supervisor.md — template for the mandatory run-identity file. The
  file is mandated by lane-policy.md but had no template — the plausible root cause of the
  exemplar run never writing one (its supervisor identity had to be recovered by transcript
  search). workflow/activation.md now also lists supervisor.md under Mandatory Artifacts.
• Wiring — workflow/activation.md (bootstrap step 10), harness README.md,
  netscript-harness SKILL (+ regenerated .claude/skills/ mirror).

Why

The exemplar run took a replan from zero to a fully-planned, GitHub-native, implementation-ready
board in one shot — empirical discovery (repo + docs + market), synthesized objectives, per-epic
design packs, adversarial + evaluator verdicts, owner-ratified one-shot filing. Owner directive:
make that pipeline repeatable for every big triage / deep-search / plan / orchestration job. This
PR freezes the stage contracts (not the exemplar's folder tree) so a fresh supervisor can
execute A→I from the doc alone.

Run artifacts

.llm/runs/harness-seed-run-profile--codify/ — supervisor.md written first (dogfooding the gate
the exemplar missed), research.md (exemplar distillation with citations), plan.md (locked
decisions LD-1..LD-8), worklog/drift/context-pack.

Gates

• validate-claude-surface.ts — all 5 checks ok (skills mirror SYNCED, 17 skills).
• Internal references in seed-run.md verified to resolve.
• fmt note: .llm/harness/**/*.md is outside the repo fmt surface (deno.json fmt.include is
  packages/plugins ts,tsx); the 29-file raw-fmt drift there is pre-existing and untouched
  (non-verdict per AGENTS.md).

Evaluation plan (this PR does not self-certify)

WSL Codex adversarial pass → fixes → OpenHands separate-session eval → owner ratification. Docs-only
change: no packages//plugins/ source touched.

🤖 Generated with Claude Code

https://claude.ai/code/session_012wKHquACkXnWPDgJYhhFjN

[rickylabs]

Slice trail (S1–S3)

Slice	Commit	Scope	Evidence
S1 run scaffolding	ebaaf935	.llm/runs/harness-seed-run-profile--codify/ — supervisor.md written first, research.md (exemplar distillation w/ citations), plan.md (LD-1..LD-8), worklog/drift/context-pack	supervisor-identity gate dogfooded; drift #2 records the exemplar's violation + root cause
S2 profile	23ccc8f6	workflow/seed-run.md (stage contracts A–I) + templates/supervisor.md (missing-template gap fix)	internal refs verified to resolve (lane-policy, resource-aggregation, tooling, labels.yml)
S3 wiring	7c5fd0d8	activation.md (step 10 + supervisor.md in Mandatory Artifacts), harness README, netscript-harness SKILL + regenerated .claude/skills/ mirror	sync-claude-skills.ts SYNCED (17 skills); validate-claude-surface.ts all 5 checks ok

Next: WSL Codex adversarial pass → fixes → OpenHands separate-session eval → owner ratification.

[rickylabs]

S4–S5: adversarial pass complete, all findings fixed

S4 (e48fcafb): WSL Codex unoriented adversarial review — fresh clone, branch-verified, brief committed to the run dir. Verdict: 0 blockers, 6 major, 2 minor, .claude/skills mirror byte-identical. Full findings: .llm/runs/harness-seed-run-profile--codify/adversarial-findings.md.

S5 (604e8ae5): all 8 findings accepted and fixed (dispositions: adversarial-triage.md):

1. Mutation boundary split into two surfaces — the run's own draft PR is the always-writable commit trail; the board (issues/epics/milestones/repo labels) is untouchable before stage H.
2. plan/ branch type registered in the netscript-pr taxonomy (seed-run only) — reconciled in the canonical home rather than contradicting it.
3. Stage B now carries the Tier-C hard rule (committed workflow.js before execution).
4. Stage F de-hardcoded from Tier D to distinct-model invariants; tier is per-run config in supervisor.md.
5. phase-registry.md scoped to multi-group runs.
6. Eval exception precisely scoped (drift Migrate shared Zod datetime codecs to Temporal #3): PLAN-EVAL skipped (owner-directed, docs-only run); separate-session OpenHands verdict retained; ships to owner for ratification — not claimed as doc-authorized.
7. Run artifacts refreshed to PR harness: codify the seed run — discovery-to-orchestration v3 run shape #471 reality.
8. FILING-LOG.md spelling unified.

Next: OpenHands separate-session eval, then owner ratification. This run does not self-certify or merge.

[rickylabs]

@openhands-agent model=openrouter/minimax/minimax-m3 output=pr-comment You are the separate-session harness evaluator for this docs-only PR. Do NOT implement fixes; produce a verdict.

Read, in order:

1. .llm/harness/workflow/lane-policy.md (the two hard invariants + supervisor-identity rule)
2. .llm/harness/workflow/run-loop.md and .llm/harness/workflow/activation.md
3. .agents/skills/netscript-pr/SKILL.md (branch/label/filing law)
4. The run dir .llm/runs/harness-seed-run-profile--codify/ (supervisor.md, research.md, plan.md LD-1..LD-8, worklog.md, drift.md, adversarial-findings.md, adversarial-triage.md)
5. The deliverables: .llm/harness/workflow/seed-run.md, .llm/harness/templates/supervisor.md, and the wiring diff vs main (git diff main...HEAD)

Evaluate:

• E1 Plan conformance: do the deliverables implement locked decisions LD-1..LD-8 without silent scope drift?
• E2 Internal consistency: does seed-run.md contradict lane-policy/run-loop/activation/netscript-pr anywhere the adversarial pass missed? Were all 8 triaged findings actually fixed in the doc text (verify each against adversarial-triage.md)?
• E3 Fresh-agent executability: could a supervisor with no chat history run stages A–I from seed-run.md alone? Name any remaining ambiguity.
• E4 Process integrity: run artifacts current, slice trail (5 commits) matches worklog, drift log honest (incl. drift Migrate shared Zod datetime codecs to Temporal #3 PLAN-EVAL exception), no packages//plugins/ source touched, no board mutation by this run.

Output a PR comment: verdict PASS or FAIL_FIX with a numbered findings list (file:line + evidence quote for each). Do not commit source changes; do not commit deno.lock churn.

[rickylabs]

OpenHands Agent — Completed

Model: openrouter/minimax/minimax-m3
Provider: OPENROUTER
Job status: failure
Trace: .llm/tmp/run/openhands/pr-471/run-28744305384-1 (committed to the PR branch)
Artifact: openhands-agent-28744305384-1

OpenHands separate-session harness evaluator — summary\n\nRun: harness-seed-run-profile--codify (PR #471, branch harness/seed-run-profile)\nMode: pr-comment, docs-only evaluation\nVerdict: PASS\n\n## Summary\n\nThis was a separate-session, read-only harness evaluator pass on PR #471. I read the workflow law (lane-policy, run-loop, activation, netscript-pr), the run dir (supervisor.md, research.md, plan.md LD-1..LD-8, worklog.md, drift.md, adversarial-findings.md, adversarial-triage.md, adversarial-brief.md, codex-thread-ids.md, context-pack.md), the deliverables (seed-run.md, templates/supervisor.md), and the wiring diff vs origin/main (5 commits: S1..S5). I verified each of the 8 triaged adversarial findings landed in the doc text, checked plan conformance for LD-1..LD-8, ran a fresh-agent executability pass, and audited process integrity (slice trail, drift log honesty, mirror parity, no packages//plugins/ source touched, no deno.lock churn, no board mutation by the run).\n\n## Changes\n\nNo code or doc changes. This run is evaluator-only. The PR comment is staged in the run-scoped reply path; the workflow owns the GitHub post.\n\n## Validation\n\n- git log origin/main..HEAD — 5 commits, matches worklog S1..S5 exactly (ebaaf93 / 23ccc8f / 7c5fd0d / e48fcaf / 604e8ae).\n- git diff --name-only origin/main...HEAD — zero packages/ and plugins/ paths; zero deno.lock changes.\n- diff -q .agents/skills/.../SKILL.md .claude/skills/.../SKILL.md — silent for both netscript-harness and netscript-pr (mirrors byte-identical).\n- Cross-reference check: each of LD-1..LD-8 has a concrete landing site in seed-run.md, templates/supervisor.md, activation.md, README.md, and/or netscript-pr/SKILL.md.\n- Triage cross-check: 8/8 adversarial findings have a corresponding line in the current doc text (F1 board-mutation invariant; F2 supervisor-identity wiring; F3 Tier-C rule; F4 unoriented/distinct-model; F5 plan/ branch type; F6 phase-registry.md conditional; F7 FILING-LOG.md + supersession; F8 drift-log honesty incl. PLAN-EVAL exception).\n- Drift log: drift.md:3 records the PLAN-EVAL exception honestly (drift #3), not laundered.\n- Board non-mutation: PR is draft, all filing deferred to Stage H, no issues.create / labels / milestones were issued by this run.\n\n## Responses to review comments or issue comments\n\nThe PR comment for this evaluator pass is in replies.json (single comment body, pr-comment mode). It carries the PASS verdict, the 4-evaluator structure (E1 plan conformance / E2 internal consistency / E3 fresh-agent executability / E4 process integrity), the 8/8 triage cross-check, the two non-blocking observations, and a Recommend merge line. Per the run contract, the workflow posts it; the evaluator does not post directly.\n\n## Remaining risks\n\n1. seed-run.md:69-70,153 — Stage B's Tier-C rule names the artifact (<run-dir>/workflows/<slice>-workflow.js) and the commit-before-run invariant, but does not point at the generator that emits workflow.js (no .llm/tools/ or tools/ cross-ref). Fresh-agent observation only; executability is not broken.\n2. seed-run.md:104,160 — FILING-LOG.md is required at H; consider linking the manifest schema (field set) in a future iteration so a fresh agent does not reinvent the column order. Observation only.\n\nNeither of these blocks this PR. Both are appropriate follow-up work for a future iteration, not for a fix-it-now edit to a docs-only PR whose scope is already locked and triaged.\n"}

Run: https://github.com/rickylabs/netscript/actions/runs/28744305384

[rickylabs]

S6 — IMPL-EVAL PASS · ready for owner ratification

Verdict: PASS ("Recommend merge") — OpenHands separate session, minimax-m3, run 28744305384. The job's failure status was the commit-back step only (known mode); the verdict is transcribed verbatim into .llm/runs/harness-seed-run-profile--codify/evaluate.md and the branch tip was verified churn-free (no trace, no deno.lock).

Evaluator confirmations: 5-commit slice trail matches worklog exactly · zero packages//plugins//deno.lock touches · skills mirrors byte-identical · LD-1..LD-8 all landed · 8/8 adversarial findings verified fixed in doc text · drift log "honest ... not laundered" · zero board mutation by this run.

Two non-blocking evaluator observations deferred to OD-1 (filing-manifest template promotion): Stage B lacks a cross-ref to the workflow.js generator tooling; Stage H lacks a linked manifest field-set schema.

Owner decision (this run does not self-certify or merge)

1. Ratify + merge: un-draft, squash-merge harness/seed-run-profile → main. Merging also ratifies drift Migrate shared Zod datetime codecs to Temporal #3: the PLAN-EVAL pass was skipped for this docs-only codification run (owner-directed); the separate-session eval and both hard invariants were retained.
2. Amend first: request changes here; the run session will fix and re-push.

Slice trail: S1 ebaaf935 · S2 23ccc8f6 · S3 7c5fd0d8 · S4 e48fcafb · S5 604e8ae5 · S6 c73bb0c4.