create-harness-vibe-coding 0.8.6 → 0.8.8

package/templates/common/.claude/skills/wf-auto/SKILL.md

@@ -1,107 +1,61 @@
----
-name: wf-auto
-description: Perpetual auto-optimization mode. Never stops — continuously improves code until 8-angle exhaustion. Adaptive checkpoints, external spark search, evidence ledger. Use for Claude /wf-auto, Codex $wf-auto, auto mode, or when the user wants unbounded self-directed optimization.
----
-
-# WF Auto — Perpetual Auto-Optimization
-
-## Load (authoritative specs)
-
-- `Harness/WF-AUTO.md` — full spec: perpetual loop, state machine, 8-angle exhaustion gate, cross-model oracle, spark candidate provider, Value Gate scoring, evidence ledger, Intent Checkpoints, anti-patterns, safety controls
-- `Harness/subagents.md` — agent roster, controller role, efficiency ladder
-- `Harness/dispatch.md` — handoff format, File claim, Concurrency group fields
-- `Harness/agent-workflow.md` — build/review/test loop, cohesion rule, completion gate
-- `.claude/skills/wf-review/SKILL.md` — cross-model invocation pattern (used by the oracle step)
-
-## Trigger & When NOT to Use
-
-- **Trigger**: Claude `/wf-auto`, Codex `$wf-auto`, `wf auto`, `auto mode`, or user wants continuous self-directed improvement
-- **Do NOT use**: user has a specific bounded task (use `/wf`), task needs maximum parallelism (use `/wf-max`), production hotfix needed urgently, codebase <100 lines
-
-## State Machine
-
-```
-auto.internal → auto.spark → auto.checkpoint → auto.exhausted → paused
-```
-
-CEO tracks state in `Harness/tasks/auto/PROGRESS.md`. Transitions are CEO-owned.
-
-## Hard Constraints
-
-1. **NEVER STOP except A-GATE.** No "task complete" early exit. The Angle Exhaustion Gate (internal + oracle + spark all empty, 3 confirm rounds) is the ONLY permitted stop.
-2. **CEO never writes production code.** CEO uses Task, Read, Grep/Glob. No Edit/Write/Bash on source files. Exception: CEO MAY write to `Harness/tasks/auto/PROGRESS.md` and `Harness/tasks/auto/PLAN.md`.
-3. **ALL sources in ONE message per cycle.** 8 angles + oracle + spark searchers. Batching is mandatory.
-4. **ONE finding per cycle.** One change, ≤3 files, ≤50 lines diff. Big ideas (>50 lines) escalate to /wf or /wf-max then return.
-5. **Two-gate review every cycle.** Spec review before code-quality review. No skipping.
-6. **A-GATE has 3 tiers.** All 8 exhausted → Cross-Model Oracle → Spark search → 2 confirmation rounds → STOP.
-7. **Value Gate is scored, not binary.** 5 dimensions (Impact, Evidence, Fit, Timing, Cost/Risk), 1-5 each. Pass: ≥18/25 AND no dimension <3.
-8. **Intent Checkpoint is adaptive.** 2→5→10 cycles. Exactly 2 questions: "Still aligned?" + "What should change?" Early on drift signals.
-9. **Evidence ledger per cycle.** Source, evidence type, expected impact, verification method, measured result, verdict. Track weak spark count.
-
-## The 8 Angles (quick reference)
-
-| # | Angle | Finds |
-|---|-------|-------|
-| 1 | Correctness | Bugs, edge cases, null safety, race conditions |
-| 2 | Performance | Slow paths, memory, algorithmic complexity |
-| 3 | Security | Injection, auth, secrets, dependency CVEs |
-| 4 | Maintainability | Clarity, DRY, coupling, naming, dead code |
-| 5 | Test Coverage | Missing tests, weak assertions, flaky tests |
-| 6 | Architecture | Boundaries, dependency direction, layer discipline |
-| 7 | UX / DX | Error messages, API ergonomics, documentation |
-| 8 | Robustness | Resilience, retry, observability, recovery |
-
-## Spark Sources (when internal + oracle empty)
-
-| # | Source | Evidence Weight |
-|---|--------|-----------------|
-| 1 | Official Docs & Advisories | HIGH |
-| 2 | Ecosystem Pulse | MEDIUM |
-| 3 | GitHub Trending (same stack) | LOW-MEDIUM |
-| 4 | Best Practices (latest) | MEDIUM |
-| 5 | Competitor/Peer Projects | LOW |
-| 6 | Real-world Issues | MEDIUM |
-| 7 | Architecture Trends | LOW |
-| 8 | Performance Benchmarks | MEDIUM |
-
-Source-quality: official docs > blog posts. Trending ≠ correct. Competitor behavior is hypothesis only. Every spark MUST cite source with URL and date.
-
-## Perpetual Loop
-
-```
-W0: SENSE (8 angles + oracle + 8 spark sources, all parallel)
-A-GATE [findings? → W1 | all empty? → oracle → spark → confirm ×2 → STOP]
-CHECKPOINT [every 2→5→10 cycles, 2 questions]
-W1: PRIORITIZE (across internal + oracle + spark)
-W2: IMPLEMENT → W3: REVIEW → W4: DEBUG → W5: VERIFY
-RECORD + EVIDENCE LEDGER → LOOP W0
-```
-
-## Cycle Recording
-
-Every cycle writes to `Harness/tasks/auto/PROGRESS.md`:
-- Cycle number, timestamp, state
-- Source (internal/oracle/spark-*), source citation
-- Finding, change description, files changed
-- Value Gate scores (if spark candidate)
-- Review result, verification evidence
-- Evidence Ledger: evidence type, expected impact, verification method, measured result, verdict
-
-## Safety
-
-- ≤3 files, ≤50 lines per cycle
-- Big ideas (>50 lines) escalate to /wf or /wf-max, then return to auto
-- Destructive changes flagged with rollback plan
-- IDLE alarm after 5 empty cycles → re-scope → A-GATE candidate
-- Spark stop: 5 failed Value Gates OR 3 weak measured impacts OR 2 repeated source families empty
-- User can interrupt at any time
-
-## Return Format
-
-- Total cycles run
-- Findings addressed per source (internal / oracle / spark)
-- Evidence ledger with measured impacts
-- Exhaustion evidence (3-round confirmation)
-- Weak spark count
-- Final codebase state
-- Residual risk assessment
+---
+name: wf-auto
+description: Perpetual adaptive auto-optimization mode. Selects probes from project evidence instead of a fixed angle count. Inherits WF acceptance gates and subagent orchestration per cycle. Use for Claude /wf-auto, Codex $wf-auto, auto mode, or unbounded self-directed optimization.
+---
+
+# WF Auto - Perpetual Auto-Optimization
+
+## Load
+
+- `Harness/WF-AUTO.md`
+- `Harness/WF-AUTO-ANGLES.md`
+- `Harness/subagents.md`
+- `Harness/dispatch.md`
+- `Harness/agent-workflow.md`
+- `.claude/skills/wf-review/SKILL.md`
+
+## Trigger
+
+- Claude `/wf-auto`
+- Codex `$wf-auto`
+- `wf auto`, `auto mode`, or a request for continuous self-directed improvement
+
+Do not use when the user gives a bounded task, requests maximum parallelism
+(`/wf-max`), needs an urgent production hotfix, or the codebase is tiny enough
+that auto scanning costs more than it helps.
+
+## Hard Rules
+
+1. Never stop except the Adaptive Coverage A-GATE: dynamic high-risk
+   obligations are covered, two different confirmation strategies are empty,
+   and unresolved uncertainty is recorded.
+2. CEO never edits production source. CEO may write only
+   `Harness/tasks/auto/PLAN.md` and `Harness/tasks/auto/PROGRESS.md`.
+3. Build a project profile each W0 cycle and dispatch only the selected probes;
+   invoke oracle and spark searchers when evidence justifies them.
+4. One accepted finding per cycle: <=3 files and <=50 changed lines. Larger
+   ideas escalate to `/wf` or `/wf-max`, then return to auto.
+5. Every accepted cycle inherits the full WF chain:
+   `Mini PRD -> AC IDs -> test/validation plan -> implementer -> verifier ->
+   cross-review -> reflector PASS -> evidence ledger -> next W0`.
+6. Review is mandatory: spec review, code-quality review, then reflector.
+7. Value Gate is scored for spark candidates: pass is >=18/25 and no dimension
+   below 3.
+8. Intent Checkpoint is adaptive: 2 -> 5 -> 10 cycles, exactly two questions.
+9. Record compact evidence per cycle; do not paste full logs or transcripts.
+
+## Loop
+
+```text
+W0: SENSE (adaptive probes + oracle + spark sources as triggered)
+A-GATE: continue, oracle, spark, confirm, or stop
+W1: PRIORITIZE one finding
+W2-W5: Mini PRD -> AC -> test/validation plan -> implementer -> verifier -> cross-review -> reflector PASS
+RECORD: evidence ledger
+LOOP: next W0
+```
+
+## Return
+
+Report cycles run, findings addressed by source, evidence ledger, exhaustion
+evidence if any, weak spark count, final state, and residual risks.

package/templates/common/.claude/skills/wf-auto-spark/SKILL.md

@@ -1,17 +1,17 @@
 ---
 name: wf-auto-spark
-description: Use for /wf-auto-spark in Claude Code, $wf-auto-spark or /skills wf-auto-spark in Codex, or perpetual inspiration mode that never stops — external spark search, long-term roadmap with staged milestones, ≤50% deviation guard.
+description: Perpetual inspiration mode for /wf-auto-spark or $wf-auto-spark. Inherits WF-AUTO/WF execution gates while using external spark search, roadmap anchoring, and <=50% deviation guard.
 ---
 
 # WF-AUTO-SPARK Adapter
 
-This skill is a thin tool adapter. The authoritative workflow lives in
-`Harness/WF-AUTO-SPARK.md`; do not duplicate or override it here.
+The authoritative workflow lives in `Harness/WF-AUTO-SPARK.md`; this adapter
+only routes and summarizes hard constraints.
 
 ## Invocation
 
-- Claude Code: use `/wf-auto-spark` or select the `wf-auto-spark` skill.
-- Codex CLI or IDE: use `$wf-auto-spark` or `/skills` then choose `wf-auto-spark`.
+- Claude Code: `/wf-auto-spark`
+- Codex: `$wf-auto-spark` or `/skills` then choose `wf-auto-spark`
 
 ## Load
 
@@ -25,15 +25,17 @@ This skill is a thin tool adapter. The authoritative workflow lives in
 
 ## Rules
 
-WF-AUTO-SPARK is perpetual inspiration mode with roadmap anchoring:
-1. **Roadmap first**: Declare North Star + staged milestones before any spark cycle.
-2. **Never auto-stop**: Only user can stop. "No sparks found" → expand search.
-3. **Deviation guard (≤50%)**: Every spark checked against North Star. Cumulative 10-cycle average ≥65%. Below → force Re-Anchor Gate.
-4. **Milestones flexible within 50%**: Can reorder/split/merge/replace, but North Star changes need user confirmation.
-5. **Value reflection every cycle**: CEO writes what was done, why it matters, deviation score, milestone progress.
-6. **Re-Anchor Gate every 10 cycles**: User confirms direction or adjusts roadmap.
-
-## Roadmap Location
-
-Active roadmap lives at `Harness/tasks/auto/SPARK-ROADMAP.md`. Created at startup.
-Per-cycle evidence at `Harness/tasks/auto/PROGRESS.md`.
+1. Roadmap first: declare North Star and staged milestones before spark cycles.
+2. Never auto-stop: only the user can stop; "no sparks found" expands search.
+3. Spark replaces discovery only. Accepted candidates still run:
+   `Mini PRD -> AC IDs -> test/validation plan -> implementer -> verifier ->
+   cross-review -> reflector PASS -> evidence ledger`.
+4. Spark searchers are read-only. Implementation requires explicit dispatch
+   packet, write set, forbidden truth files, AC IDs, and verification commands.
+5. Deviation guard: each spark must align >=50% with North Star; rolling
+   10-cycle average below 65% forces Re-Anchor Gate.
+6. Value reflection is required every cycle: source, why it matters, deviation,
+   evidence, and milestone progress.
+
+Active roadmap: `Harness/tasks/auto/SPARK-ROADMAP.md`.
+Per-cycle evidence: `Harness/tasks/auto/PROGRESS.md`.

package/templates/common/.claude/skills/wf-max/SKILL.md

@@ -1,18 +1,17 @@
 ---
 name: wf-max
-description: Use for /wf-max in Claude Code, $wf-max or /skills wf-max in Codex, or maximum-parallelism Harness work with CEO to manager to worker decomposition.
+description: Use for /wf-max in Claude Code, $wf-max or /skills wf-max in Codex, or maximum-parallelism Harness work with WF strict-superset gates and CEO -> manager -> worker decomposition.
 ---
 
 # WF-MAX Adapter
 
-This skill is a thin tool adapter. The authoritative workflow lives in
-`Harness/WF-MAX.md`; do not duplicate or override it here.
+The authoritative workflow lives in `Harness/WF-MAX.md`; this adapter only
+routes and summarizes hard constraints.
 
 ## Invocation
 
-- Claude Code: use `/wf-max [task]` or select the `wf-max` skill.
-- Codex CLI or IDE: use `$wf-max` or `/skills` then choose `wf-max`.
-- Codex app may also list enabled skills in the `/` menu.
+- Claude Code: `/wf-max [task]`
+- Codex: `$wf-max` or `/skills` then choose `wf-max`
 
 ## Load
 
@@ -26,18 +25,38 @@ This skill is a thin tool adapter. The authoritative workflow lives in
 
 ## Rules
 
-WF-MAX is a three-layer architecture:
-1. Global mode (`wf-max`)
-2. Agent role (`ceo` | `manager` | `worker` | `reviewer`)
-3. Dispatch permission (`writeSet`, `forbidden`, `verification`)
-
-- Top-level orchestrator is CEO (reads, plans, dispatches). Delegated Workers follow dispatch packet with explicit writeSet. Global mode ≠ every agent is CEO.
-- CEO: never edit source files directly. Spawn Workers with writeSet.
-- Worker: edit only files in dispatch writeSet. Outside writeSet → blocked.
-- Manager: scope, review, coordinate. No source edits by default.
-- Reviewer: read and report only. No edits.
-- Use the D-GATE in `Harness/WF-MAX.md` before any implementation wave:
-  dispatch table, self-audit, disjoint file claims, and reviewer plan.
-- Use real subagents when the runtime supports them; otherwise record a
-  bounded-pass fallback in `Harness/tasks/<task-id>/PLAN.md`.
-- Keep `Harness/tasks/<task-id>/PROGRESS.md#Heartbeat` current.
+WF-MAX is a WF strict superset: every WF role, gate, and acceptance rule still
+applies, then execution expands through:
+
+1. Global mode: `wf-max`
+2. Agent role: `ceo | manager | worker | reviewer | verifier | reflector`
+3. Dispatch permission: `writeSet`, `forbidden`, `verification`
+
+- CEO reads, plans, dispatches, synthesizes, and writes task state only. CEO
+  never edits production source.
+- Workers edit only the dispatch `writeSet`; outside write set is blocked.
+- Managers coordinate and synthesize. Reviewers read/report only.
+- D-GATE is mandatory before implementation waves: dispatch table, AC IDs,
+  disjoint file claims, self-audit, and reviewer plan.
+- Final acceptance requires verifier evidence, cross-review, and reflector PASS.
+
+## Fan-Out Discipline
+
+- Use as many useful subagents as the runtime safely allows.
+- Codex capacity may be configured through official `agents.max_threads` and
+  `agents.max_depth`; generated Harness config defaults to
+  `.codex/config.toml` with `max_threads = 12` and `max_depth = 1`.
+- Close completed agents before declaring the pool exhausted.
+- If Codex remains bottlenecked, ask the user before raising
+  `agents.max_threads` above the scaffold default. Do not silently edit project
+  or global Codex config.
+- Keep `agents.max_depth = 1` unless the user explicitly approves recursive
+  delegation.
+- If the current runtime is exhausted, overflow to the other CLI with explicit
+  dispatch packets: Codex -> `claude -p`; Claude -> available Codex CLI such as
+  `codex exec`.
+- Do not rely on undocumented config, environment variables, forked/derived
+  conversations, Codex++, local patches, or third-party forks as stable ways to
+  remove subagent limits.
+- Fall back to bounded role passes only when native subagents and cross-CLI
+  overflow are unavailable; record the fallback in the task PLAN.

package/templates/common/.claude/skills/wf-readme/SKILL.md

@@ -1,49 +1,49 @@
----
-name: wf-readme
-description: Use when a project README already exists and the user asks to preserve, merge, modernize, optimize, or clarify repository documentation during harness install or documentation work.
----
-
-# README Optimizer
-
-Improve `README.md` without breaking project-owned public docs.
-
-## Load
-
-- root `README.md`
-- package files and scripts (`package.json`, `pyproject.toml`, `go.mod`, etc.)
-- CI files when present
-- `Harness/PROGRESS.md`
-- `Harness/tasks/<task-id>/PLAN.md` when available
-- `Harness/architecture.md` only when an architecture summary or diagram is requested
-
-## Mode
-
-Ask the user which mode they approve when the existing README is meaningful:
-
-| Mode | Use when | Allowed edit |
-| --- | --- | --- |
-| Preserve + append | default for existing projects | Add only a compact Development, Test, Build, Git, or Harness section |
-| Structure pass | README is stale, hard to scan, or missing operational docs | Reorganize with headings, tables, command blocks, and links while preserving facts |
-| Full rewrite | user explicitly wants a polished public README | Rewrite after approval; keep claims source-backed |
-
-If unanswered, use Preserve + append.
-
-## Rules
-
-- Preserve existing product, package, API, and public-facing content unless the user approves a rewrite.
-- Do not invent features, benchmarks, roadmap, support policy, badges, install commands, or CI status.
-- Use tables for command matrices, environment variables, endpoints, and deployment notes when facts are known.
-- Use Mermaid or ASCII architecture diagrams only when the structure is observed or approved; label uncertain diagrams as proposed.
-- Keep detailed architecture in `Harness/architecture.md`; README may link to it or show a short overview.
-- Keep agent rules in `CLAUDE.md`/`AGENTS.md`, not README.
-- Record the chosen mode and any skipped README improvements in `Harness/tasks/<task-id>/PLAN.md` when available.
-
-## Output
-
-Before broad edits, return:
-
-1. chosen mode
-2. sections to preserve
-3. sections to add or reorganize
-4. facts still unknown
-5. verification command or manual review step
+---
+name: wf-readme
+description: Use when a project README already exists and the user asks to preserve, merge, modernize, optimize, or clarify repository documentation during harness install or documentation work.
+---
+
+# README Optimizer
+
+Improve `README.md` without breaking project-owned public docs.
+
+## Load
+
+- root `README.md`
+- package files and scripts (`package.json`, `pyproject.toml`, `go.mod`, etc.)
+- CI files when present
+- `Harness/PROGRESS.md`
+- `Harness/tasks/<task-id>/PLAN.md` when available
+- `Harness/architecture.md` only when an architecture summary or diagram is requested
+
+## Mode
+
+Ask the user which mode they approve when the existing README is meaningful:
+
+| Mode | Use when | Allowed edit |
+| --- | --- | --- |
+| Preserve + append | default for existing projects | Add only a compact Development, Test, Build, Git, or Harness section |
+| Structure pass | README is stale, hard to scan, or missing operational docs | Reorganize with headings, tables, command blocks, and links while preserving facts |
+| Full rewrite | user explicitly wants a polished public README | Rewrite after approval; keep claims source-backed |
+
+If unanswered, use Preserve + append.
+
+## Rules
+
+- Preserve existing product, package, API, and public-facing content unless the user approves a rewrite.
+- Do not invent features, benchmarks, roadmap, support policy, badges, install commands, or CI status.
+- Use tables for command matrices, environment variables, endpoints, and deployment notes when facts are known.
+- Use Mermaid or ASCII architecture diagrams only when the structure is observed or approved; label uncertain diagrams as proposed.
+- Keep detailed architecture in `Harness/architecture.md`; README may link to it or show a short overview.
+- Keep agent rules in `CLAUDE.md`/`AGENTS.md`, not README.
+- Record the chosen mode and any skipped README improvements in `Harness/tasks/<task-id>/PLAN.md` when available.
+
+## Output
+
+Before broad edits, return:
+
+1. chosen mode
+2. sections to preserve
+3. sections to add or reorganize
+4. facts still unknown
+5. verification command or manual review step

package/templates/common/.claude/skills/wf-remove/SKILL.md

@@ -1,10 +1,10 @@
----
-name: wf-remove
-description: Use for /wf-remove in Claude Code, $wf-remove or /skills wf-remove in Codex, or any request to uninstall Harness safely.
----
-
-# WF Remove Adapter
-
+---
+name: wf-remove
+description: Use for /wf-remove in Claude Code, $wf-remove or /skills wf-remove in Codex, or any request to uninstall Harness safely.
+---
+
+# WF Remove Adapter
+
 ## Invocation
 
 - Claude Code: use `/wf-remove` or select the `wf-remove` skill.

package/templates/common/.claude/skills/wf-update/SKILL.md

@@ -23,13 +23,18 @@ description: Use for /wf-update in Claude Code, $wf-update or /skills wf-update
    `agent` block as the action plan.
 2. Preserve all PRESERVE files. Never overwrite user task, memory, research,
    README, package, or architecture files.
-3. If `agent.safeApplyCommand` is present, run it to apply SAFE and NEW files
-   before spending AI time on conflicts. Default command:
+3. If `agent.safeApplyCommand` is present, run it to apply SAFE, NEW, and
+   adopted metadata-only files before spending AI time on conflicts. Default command:
    `node Harness/scripts/wf-update-check.mjs --apply-safe`.
 4. For every `agent.aiMergeRequired` entry, compare the local file with
    `templateHint` or `remoteUrl`, then choose merge, keep-local, or
-   overwrite-from-template. Ask the user only when the intent is ambiguous.
-5. Use strict `--apply` only when the JSON plan has zero conflicts.
+   overwrite-from-template. Record the decision through the script with
+   `--accept-local <file>`, `--accept-merged <file>`, or
+   `--accept-template <file>`; do not hand-edit `Harness/.harness-version`.
+   Ask the user only when the intent is ambiguous.
+5. Run `node Harness/scripts/wf-update-check.mjs --finalize` after all
+   conflicts have script-recorded decisions. Use strict `--apply` only when the
+   JSON plan has zero conflicts.
 6. After update, run the validator and then scan-clean.
 
 ## Return

package/templates/common/.codex/config.toml

@@ -0,0 +1,5 @@
+[agents]
+# Harness WF-MAX default: higher than Codex's unset default of 6.
+# Keep depth at 1 unless the user explicitly approves recursive delegation.
+max_threads = 12
+max_depth = 1