Suggest Proactive

Corgi proactive suggest

A scheduled push wrapper around suggest: on a cadence (owned by /schedule), runs the suggest ranking, takes the top idea, dedupes it against open tickets + recently-dismissed ones, then either proposes it and asks (default) or — only if the workspace opted in — files exactly one draft tracker ticket. Reuses suggest for ranking and tracker for the write gate — does NOT duplicate either. Default is propose-and-ask; the opt-in files one draft per run, rate-limited + deduped. Inspired by the "proactive engineer" idea: push, not pull.

Guardrails (non-negotiable)

• Reuse, don't duplicate. Ranking + evidence + measurable-outcome rules are owned by suggest; issue creation goes through the tracker write gate + MCP preflight. Never re-implement either, never reinvent scheduling — /schedule owns it.
• One ticket per run, max. Per-week cap (default 1, hard ceiling 3) enforced via corgi_services/suggest-history.json (corgi suggest-history check/record).
• Default is propose+ask. Auto-file requires the explicit suggest.autoFileDrafts flag in ~/.corgi/config.yml and is draft-only: never assign, never move past draft/backlog, never trigger stories/a build, never open a PR.
• Dedupe before filing: open tickets, the state file (filed always + dismissed within cooldown + proposed within cooldown), and [[workspace-memory]] (loose, best-effort — see memory skill).
• Headless = never silently file in propose mode. No human to confirm → record proposed, notify, exit. The cron fires while the REPL is idle — treat it as headless.
• Idempotent. Match on slug/title; update or skip, never create a duplicate.
• Print the active mode at the top of every run so it's never a surprise.
• No AI attribution. Everything posted or committed — ticket comments, PR/MR bodies, commit messages, store listings/notes — reads as if a human wrote it. Never add "generated by AI / Claude / corgi", a co-authored-by or AI trailer, a bot signature, or a robot-emoji badge to any outward text.

Phase 0 — Resolve workspace + mode + state

1. Workspace. Resolve the path from $ARGUMENTS (absolute) or cwd. cd there. Require corgi-compose.yml (ls corgi-compose.yml *.corgi-compose.yml) — absent → stop with the "open the workspace folder" message. The cron fires with no implied cwd, so the schedule prompt always names the absolute path (Phase 5).
2. Mode. Read corgi suggest-history config --json → {autoFileDrafts, maxPerWeek} (fallback: read ~/.corgi/config.yml directly; absent → propose, cap 1). Print the active mode line, e.g. proactive suggest · mode=propose · cap=1/week.
3. State. Load history via corgi suggest-history list --json (fallback: read corgi_services/suggest-history.json; absent → empty). The helper is optional at runtime — if the binary predates it, read/append the JSON directly with the same shape.

Phase 1 — Rank (reuse suggest)

• Invoke the suggest skill flow (its Phases 0–3) scoped to this workspace, both lenses, to produce the ranked shortlist. Do NOT re-spec everything — only the chosen survivor gets specced (Phase 3), and only after it survives dedupe.
• State explicitly to yourself: ranking + evidence + measurable-outcome rules are owned by suggest; this skill only consumes the ranked list.

Phase 2 — Dedupe + rate-limit

• Derive a slug per candidate deterministically from its title (lowercase, non-alphanumeric → -, collapse repeats, trim) — same idea ⇒ stable key across runs. (corgi suggest-history record and the helper use the same Slugify.)
• Top candidate down, for each: corgi suggest-history check --slug <slug> --json → { "skip": bool, "reason": "filed|dismissed|proposed|rate-limit", "slug": ... }. Also fuzzy-match the candidate title/slug against open tickets via tracker Phase-0/1 (the corgi-suggest label set + title match) — a hit there is open-ticket, skip. And consult [[workspace-memory]] (below): a matching "dismissed suggestion" / "won't build" slug is treated like a dismissed state-file entry.
• The first surviving candidate wins. If none survive, or check returns reason=rate-limit (week cap already used) → no-op: record a skipped audit entry (corgi suggest-history record --slug <slug> --status skipped), report what was deduped and why, go to Phase 5. Don't file anything.

Phase 3 — Decide: propose vs auto-file-draft

• propose (default): spec the survivor — reuse suggest's Phase 4 spec shape → docs/suggestions/<slug>.md — and present it. Then:
  • Human present → ask "file this as a tracker story?" (identical to interactive suggest Phase 5). Yes → Phase 4 (the tracker write gate). No → record dismissed so the cron won't re-propose it within the cooldown; stop.
  • Headless (cron, idle REPL, no one to answer) → do not file. Record proposed (corgi suggest-history record --slug <slug> --status proposed --title <…> --lens <…>), fire a PushNotification if available, stop. Surface it next time a human is present.
• auto-file-drafts (explicit opt-in): go straight to Phase 4 to file one draft.

Phase 4 — File (reuse tracker write gate) OR record-and-report

• Preflight the matching tracker MCP (Linear mcp__linear-server__* vs Jira mcp__atlassian__*, detected as in suggest Phase 5 / tracker Phase 0). Not connected → do NOT silently file: keep docs/suggestions/<slug>.md + a paste-ready body + the tracker's new-issue URL, record proposed, stop.
• Connected → create the issue in the draft/triage/backlog-equivalent state, labels corgi-suggest + draft, no assignee, link the spec. Reuse the same tracker write path — do not hand-roll the MCP call shape. Then record it: corgi suggest-history record --slug <slug> --status filed --ticket <KEY> --title <…> --lens <…>.
• Never offer "implement it now?" here. That's the interactive suggest/stories path — proactive stays hands-off: never assign, never move past draft, never build, never open a PR.

Phase 5 — Arm/maintain the schedule

• Point at references/schedule-config.md for the literal CronCreate args.
• If the user asked to set it up: emit the config (weekly, off-:00 minute, durable: true, the absolute workspace path in the prompt) and warn about the 7-day auto-expire (CronCreate recurring jobs fire a final time then delete after 7 days) — offer to re-arm, and offer a --once next-Monday one-shot as the conservative default for first-time users.
• Disarm: CronList → CronDelete by id (see the reference).

[[workspace-memory]] integration (loose)

If a workspace-memory store is present (.corgi/memory/, the memory skill), do a best-effort read for entries tagged "dismissed suggestion" / "won't build" and treat a matching slug the same as a dismissed state-file entry — so an idea the user explicitly rejected in conversation isn't re-proposed by the cron. Advisory only: degrades silently when no memory store exists, and the state file remains the source of truth. Match on slug/title text; never block on the memory API. Don't duplicate the memory convention here — see the memory skill.

Scenarios

• First-time setup → prefer the --once next-Monday one-shot; explain the 7-day expire before arming anything recurring.
• Every candidate deduped / week cap hit → a clean no-op with a skipped audit entry, so gain/suggest-history list shows the job ran and why nothing filed.
• No tracker MCP → never guess a tracker; record proposed and report the new-issue URL.

Source: Andriiklymiuk/corgi — distributed by TomeVault.