feat(agent-tasks): daily-ticket-status-drafts A/B harness (v0)

[alycda]

First Cowork scheduled task migrated to a runtime-agnostic harness, for an A/B quality comparison against the still-enabled Cowork incumbent (through 2026-09-01). Plan/requirements/ideation live in the private migration corpus (~/agentic-migration/docs/).

What

• tools/agents/tasks/daily-ticket-status-drafts/ — task bundle: README, codex --output-schema contract (drafts XOR evidence-carrying skip, per-ticket dispositions, complete attestation), and the PreToolUse hook gate (hooks/never-post-linear.json: probe + deny-hook, hash-trusted install).
• tools/agents/bin/ticket-drafts-run — drafting arm wrapper. Default-red: green requires pending drafts or an evidence-carrying skip. Distinct red taxonomy (auth | connector | rate-limit | timeout | malformed-output | incomplete | evidence-free-skip | placeholder-secret) so arm-unavailable days never count as quality data. Codex runs read-only sandbox; the wrapper owns all outbox writes; records named <date>-<ticket>-<runid>.json with supersede-by-state-transition (never overwrite). Sentinel refusal prevents drafting from the committed placeholder prompt.
• tools/agents/bin/ticket-drafts-review — the only posting path (holds the Linear key). Per-draft approve / edit-then-approve / deny; approval freezes final text; posting is idempotent (attempt marker + frozen-body match on retry; editing a failed post revokes approval). Lazy calendar-date expiry with post_failed/approved exempt and resurfaced first. Cross-arm duplicate warning (24h window vs UTC-safe timestamps). Every decision → decisions.jsonl (the A/B dataset). Key passed to curl via fd, never argv.
• home-manager/modules/tools/agent-tasks.nix — work-scoped, mkIf isDarwin-guarded (work.nix also backs the alyssa@work-dev Linux devcontainer — verified the guard: Linux config gains no secrets). Deploys scripts via writeShellApplication (shellcheck at build; jq/curl/rage/coreutils pinned) and the bundle to ~/.agents/tasks/.
• justfile: ticket-drafts, ticket-drafts-one TICKET (injection-safe via quote()), ticket-drafts-review. Checkout scripts outrank stale PATH copies; everything works pre-activation via rage-direct secret fallback.

Verification

• just ci clean; both script derivations build (shellcheck green).
• Full darwin-rebuild switch on the dotfiles-ci tart VM: exit 0; scripts on PATH; bundle deployed; both secrets decrypt at 0400; expected-red --dry-run (auth, codex absent) writes its run record — default-red proven.
• Offline scenario tests: sentinel refusal, invalid-ticket red, stale-pending spike, post_failed/approved midnight exemption + resurface, dry-run spikes excluded from decisions.jsonl, injection attempt reds at the wrapper.
• Code review (8-angle find + verify) applied 11 fixes pre-merge, incl. a bash ${2:-{}} expansion bug that would have broken every Linear GraphQL call, and GNU-mktemp incompatibility on the nix-deployed path.

Before first real run (user actions)

1. agenix -e secrets/personal/ticket-drafts-prompt.age — real prompt (adapted Cowork SKILL.md + comment-patterns.md).
2. Mint a Linear personal API key → agenix -e secrets/personal/linear-api-key.age.
3. Hook gate probe per hooks/never-post-linear.json, then echo true|false > ~/.agents/tasks/daily-ticket-status-drafts/hook-gate-state.

Known residuals

• Linear GraphQL field shapes (issue(id:) accepting identifiers, user.isMe, commentCreate client-ID support) are unverifiable until the key exists — ticket-drafts-review isolates them in one documented integration point.
• Hook-visibility of account-level connectors is undetermined until the probe runs; if invisible, the frozen-snapshot increment gets promoted per the plan.
• Codex auth detection = codex login status exit code (probed on 0.143.0).

🤖 Generated with Claude Code