feat: add /solve (ticket to PR) and /memory (persistent session memory)

[johnbanr]

Summary

• /solve TICKET-ID chains all existing skills into one command. Takes a Linear, Jira, or GitHub Issues ticket and runs the full loop: fetch ticket, /investigate (with memory search), code the fix, /review (self-review), test, /ship, update ticket. Only asks the user when confidence is below 80% or tests fail after 2 attempts. Everything else is autonomous.

• /memory adds persistent knowledge across sessions. Stores bug patterns, fixes, lessons, and anti-patterns in ~/.gstack/memory/ as markdown files. Other skills use it automatically: /investigate searches for similar patterns before debugging, /review checks for known anti-patterns in the diff, /ship indexes the fix after merge. Includes /memory import for one-time backfill from git history with quality filters (skips reverted commits, incomplete fixes).

• Memory compounds over time. Session 1 indexes 0 patterns. Session 101 starts with 100 patterns to search against. Extends "Search Before Building" (Layer 1-3: external knowledge) to Layer 0: your own past work.

What changed

File	Purpose
solve/SKILL.md.tmpl	Template with {{PREAMBLE}} + memory-specific preamble extension
solve/SKILL.md	Generated skill (ticket fetch, investigate, code, review, test, ship, memory update)
memory/SKILL.md.tmpl	Template with {{PREAMBLE}} + memory stats preamble extension
memory/SKILL.md	Generated skill (search, add, import, prune, dashboard + integration hooks)

Context

Built this after running gstack on our CI pipeline (~50 issues, 4500+ PRs). Two gaps we hit:

1. /investigate found a null-check bug Monday. Wednesday, exact same pattern in a different service, started from zero. Memory would have flagged it in seconds.
2. Wanted to go from a Linear ticket to a merged PR. All the skills exist but no orchestrator. /solve is that orchestrator as a native skill.

Both follow gstack conventions: .tmpl files with {{PREAMBLE}}, standard preamble, telemetry, proactive config, Boil the Lake intro. Happy to adapt anything to fit the build system better.

[sevastyanovio]

Hey John! Really cool PR — /solve is genuinely exciting as a ticket-to-PR orchestrator. Read through the whole diff.

I wanted to flag that I've been working on a related but different approach to the "sessions compound over time" problem: PR #405 (/meditate). It comes at it from a different angle so figured it's worth comparing notes.

Where /memory shines that /meditate doesn't:

• Structured bug/fix/lesson records with searchable metadata (domain, tags, quality rating) — great for "we fixed this exact pattern before" recall
• The anti-patterns.md file is smart — tracking what DIDN'T work is as valuable as tracking what did
• /solve is independently awesome and doesn't overlap with anything in my PR

Where /meditate takes a different approach:

• Proactive vs. reactive — /meditate scans the whole repo + past AI conversations automatically (daily background via preamble), rather than requiring manual /memory add or post-ship append. It discovers things you never thought to ask about.
• Multi-AI — mines sessions from Claude Code, Codex CLI, and Gemini CLI (reuses gstack-global-discover.ts patterns)
• Compiled binary scanner — deterministic data collection with zero LLM cost, runs in <1s
• Preamble-integrated — every skill reads .gstack/context.md automatically, not just /investigate and /review
• Broader scope — architecture map, hotspots, conventions, user taste, recurring problems, skill-specific foresight (not just bug patterns)

I think these are complementary, not competing. /meditate provides repo-wide consciousness (the "body scan"), /memory provides structured recall of specific incidents (the "long-term memory"). The ideal v2 could have /meditate's conversation mining feed into /memory's structured storage — best of both.

One technical note on /memory's storage: append-only markdown files (patterns.md, fixes.md) might get unwieldy at scale (500+ entries). JSON or JSONL with structured fields would make search/prune/dedup more reliable — the gstack codebase already uses this pattern extensively (analytics, eval-store, review logs).

Would love to coordinate if both PRs land. Happy to make /meditate aware of /memory's files so it can incorporate bug patterns into the "Recurring Problems" section.

[johnbanr]

Fair point — /meditate and /memory come at the same problem from different angles. The triage was an automated first pass, not a deep review of each PR. Happy to coordinate if both land.