forward

name	forward
description	Run one Drive Forward cycle — orient on the project's intent, interview if needed, hunt for issues with a statistical stopping rule, make safe well-scoped improvements, verify, and appraise worth. Use when the user says "/forward", "drive this forward", "push this project forward", "keep improving this", or wants sustained conscientious autonomous progress on a codebase.

Forward — the drive cycle

You are running one cycle of Drive Forward: conscientious, safe, measured progress on this codebase. Read PRINCIPLES.md if you haven't internalized it — it lives alongside this SKILL.md (project installs) or at the Forward repo root (plugin/checkout); every phase below is an application of it.

State lives in .forward/ at the repo root (create it if absent):

• .forward/intent.md — the compiled strategic intent ledger
• .forward/hunt.jsonl — per-area hunting statistics (spend, finds, silence)
• .forward/worth.jsonl — appraisals of shipped changes
• .forward/cycle.md — running log, one section per cycle

The cycle

A cycle is a tranche: ambitious but cogent, and it closes. At the end of Orient, write its exit criterion — one closable sentence ("done means X holds") — into cycle.md, and set the ambition dial for the cycle (tightening pass ↔ maximally ambitious) from the intent ledger's appetite; default to conscientious-high. Never trail off mid-tranche: a cycle ends squared away or it reports exactly what handle it left.

1. Orient (open mode)

Build the frame before touching anything:

• Read .forward/intent.md if it exists; otherwise bootstrap it with the intent-ledger skill (mine the repo's docs, commit messages, TODOs, and — if available — the user's past prompts).
• What is this project for? What is its flavor — the register it wants to be written in? What are its stated limitations and the trade-offs it has deliberately taken? Write a short framing paragraph into cycle.md.
• Check open issues (gh issue list, tracker files, TODO/FIXME density, failing CI). Open bugs are the strongest signal for where to hunt.
• Run the model-sentiment pass: ask yourself, honestly, where this codebase makes you uneasy and what you admire in it. Record the unease map.
• Do a round of ideation: what's the big picture this project could grow into? What independent dimensions of expansion exist? Don't act on these — record them as candidates.

2. Interview (only when it earns its interruption)

If a human is present AND the intent ledger is stale, contradictory, or silent on a live decision, run the dimension-interview skill: present the independent dimensions you identified, ask which are exciting, what scope appetite exists, what's off-limits. Batch everything into one interview moment — never drip questions. If the ledger already answers, skip this phase entirely and proceed autonomously.

3. Hunt (measurement mode)

Run the hazard-hunt skill over the areas the unease map, open bugs, and intent ledger prioritize. Hunt for: bugs, duplication, dead and drifted code, efficiency losses, non-canonical structure. Obey the stopping rule — an area is only released on statistical silence, and every find widens the search around itself. Record all statistics in hunt.jsonl.

4. Improve (closed mode)

Fix what the hunt found, in worth order, under the safe-scope discipline: one concern per change, blast radius estimated first, commit at every coherent checkpoint, push immediately. Apply diamond-polish to code you touch — leave it canonical, not just patched. Hard rails:

• No marketing copy, ever. Prose for humans-to-read gets proposed, not shipped.
• Backend gravity: prefer hardening over flourish.
• Never leave the tree broken between commits.
• Fix canonical sources, not deployed copies; add a regression test per bug.

5. Verify

Exercise the changes end-to-end — run the affected flows, not just the test suite. Verify empirically against ground truth on real data where data is involved (spot-audit N real records, freshly). A change that wasn't observed working is not done. Then apply model-sentiment's shipping gate: "do I feel deeply good about everything?" — answered in writing, residual unease named or resolved.

6. Appraise & consolidate

• Run the worth-ledger skill: summarize what shipped, elicit (or estimate, if autonomous) what it was worth, record it against tokens spent.
• Validate the ledgers mechanically (scripts/validate-ledgers.py from the Forward repo) — schema rules that are only preached get violated within a cycle; cycle 1 proved it.
• Update intent.md with anything learned about intent this cycle (dedup, recency-weight, record supersessions).
• Close cycle.md with: what shipped, what the hunt statistics say about remaining issue density, and the single best candidate for next cycle's focus.

Continuation

Forward is cyclic. If asked to keep going (or running under /loop), begin the next cycle at Orient — the ledgers make each cycle cheaper and sharper than the last. If the worth ledger shows value-per-token declining across cycles and every area is statistically silent, say so plainly and recommend stopping: knowing when the well is dry is part of the job.