Asset validation

Asset Validation

Validate reusable assets through review, acceptance design, real CLI execution, evidence capture, and cleanup. You are the observer. Asset-under-test runs as real CLI in tmux, never as subagent.

HARD CONSTRAINTS

• MUST run the asset-under-test as real CLI in tmux; DO NOT use a subagent as the asset-under-test.
• DO NOT query SQLite directly. All state reads/writes go through scripts/acc.py.
• DO NOT create or update memory, global notes, caches, or host config unless the user explicitly asks.
• NEVER print settings files or secret env values; report paths only.
• NEVER manufacture a pass by doing the asset-under-test's work in the observer.
• NEVER glob-clean /tmp/acc-* or kill broad tmux state; rely on acc finalize / acc cleanup --round.
• DO NOT run acc --help for discovery; contracts here and in references are authoritative.
• MUST feed tasks via acc feed-task --round or acc profile run-task; only acc finalize leaves running.
• Plugin assets MUST be staged via sandbox settings/--plugin-dir; DO NOT use real or symlinked HOME skill roots as install evidence.

Convergence And Task Design

Before tasks or verdicts, use references/convergence-and-task-design.md: observer-owned verdicts, post-fix clean-pass, capability profiling, progressive ladders. A tiny task is only smoke for complex assets.

Command Entry

Derive <skill_dir> from this skill folder, not a hard-coded global install path. The ACC resolver is defined in references/unattended-execution.md (staged copy -> repo checkout -> loaded skill dir); use that resolver verbatim.

All state reads/writes go through scripts/acc.py. If a read is missing, add a narrow acc command first.

Thin script entry. Keep scripts/acc.py thin (args, dispatch, JSON, error shaping). Put durable capabilities in focused submodules: DB, env prep, plugin staging, tmux observation, cleanup.

State And Config

State lives in ~/.acceptance/state.sqlite3; ACCEPTANCE_HOME overrides for tests or isolation. Select CLI at start/launch: ask claude/codex in attended mode, default claude when absent.

acc start idempotently prepares sandbox workdir, isolated DB root, ACCEPTANCE_TMPDIR, runtime roots, and Claude settings. Use returned isolation_env.

Entry Routing

• @asset plus 评估 / 迭代 / 验收 / acceptance / evaluate / validate: full pipeline.
• @asset plus review: review only; stop after review-and-fix.
• @asset with no verb: ask for full pipeline / acceptance only / review only.
• Multiple unfinished acceptances: acc accept list --asset <name> --status draft and ask which to continue.

Unattended Mode

When the user is away, sleeping, or asks for automatic completion, treat confirmations as pre-authorized unless the action changes destructive scope or touches assets outside the asset-under-test.

Use references/unattended-execution.md before unattended rounds (command spine, temp-file rules, prompt shape, tmux capture, cleanup).

Phase Checklist

1. Classify the asset; first write with acc bootstrap (acc asset add + acc accept new is the two-step equivalent). Attended: confirm type/purpose first; unattended: record pre-authorized confirmation.
2. Write capability profile and progressive observer-scored tasks. See convergence reference.
3. Review/fix major logic only in the asset-under-test. See references/review-and-fix.md.
4. acc accept update: strategy, acceptance prompt, criteria, task prompts, ladder, fixture. Flat task JSON: {"t1": "body to send"}.
5. Pick scheduling mode with reason: stop-loss / collect-first / hybrid.
6. Launch observe loop only after prompt, criteria, tasks, fixture, cleanup plan, and ladder coverage are explicit.
7. After each round: finalize/clean, fix defects, re-run from acc start until fresh clean PASS or blocked.
8. Before feeding tasks, confirm each is a black-box stimulus; observer scores from transcripts, commands, files, processes, cleanup.

Observe Loop

Real CLI spine: acc start --cli <claude|codex> -> acc launch --round <round_id> --cli <claude|codex> -> acc feed-task --round <round_id> --task t1 -> bounded wait -> acc capture -> acc record -> independent re-verification -> acc finding as needed -> acc finalize. Omit --cli to default claude. Default acc finalize cleans sandbox, nested sandboxes, plugin staging, and round tmux. Use --keep-sandbox only for local debug, then acc cleanup --round <round_id> before return.

Only acc finalize changes a round out of running; DO NOT poll the DB for that transition.

After finalize, inspect verdict and evidence. On FAIL/CONDITIONAL or incomplete cleanup/evidence, keep working per the convergence reference.

Side Effects

Write only the acceptance DB, round sandbox, fixture/evidence paths, and authorized asset-under-test fixes. DO NOT create or update memory, global notes, caches, or host config unless the user explicitly asks.

Gotchas

• Correct-looking answer without skill/agent/plugin/command/transcript evidence is a bypass.
• Direct sqlite3 .../state.sqlite3 is a bypass; use acc round list, acc history, acc show, or add an acc read.
• NEVER print settings files or secret env values; report paths only. acc capture/acc record and every structured acc read redact known secret keys and high-confidence bare tokens. The rig also sanitizes evidence persisted by older versions before returning round list or history; add a regression whenever a new token shape is observed.
• Bash variables do not persist across tool calls; resolve ACC each Bash or store in "$WORK/.acc-path".
• Self-review is not a negative trigger test. Observer MUST watch behavior, not ask the asset to grade its own trigger.
• One-line toy task is smoke only; not final verdict for complex orchestration/recovery/workflow assets.
• NEVER manufacture a pass by doing the asset-under-test's work in the observer.
• NEVER glob-clean /tmp/acc-* or kill broad tmux; rely on acc finalize auto-cleanup, or acc cleanup --round for kept/debug/orphan rounds.
• Fixed scratch paths (/tmp/acc-toy, /tmp/acc-work-path.txt, /tmp/.<name>_marker) collide. Persist paths only under the round sandbox or its ACCEPTANCE_TMPDIR, e.g. "$WORK/.workpath".

Asset Strategies

Before final verdict, apply references/asset-strategies/common.md plus the per-type file. Skill assets need script-entry assessment and, for self-validation/validation-harness skills, the outer/nested round boundary.

• references/asset-strategies/common.md
• references/asset-strategies/skill.md
• references/asset-strategies/plugin.md
• references/asset-strategies/rule.md
• references/asset-strategies/agent.md