feat: agent loop, tool layer, and skills runner on rig

[aksOps]

Implements the agent loop, tool layer, and skills runner from #68, built on the new rig library (v0.1.0) extracted from kb's AI substrate.

What changed

• Tool layer (internal/server/tools.go): propose_card (validated via the existing validateDraft path; the cap is enforced by a collector whose limit the model can see), find_similar (FTS via SearchSimilar), fetch_link (dedicated SSRF-guarded client with its own KB_LINK_ALLOW_PRIVATE policy, 64 KiB cap, text/json only), and list_tasks / get_task / update_task mirrors of the MCP tools.
• Skills (internal/server/skills.go): embedded defaults under internal/server/skills/*.md, overridable per name from KB_DATA/skills. Names and descriptions are advertised in the system prompt; bodies load on demand via load_skill or are force-injected on explicit invocation.
• POST /api/ai/run-skill (internal/server/runner.go): generic runner endpoint returning {cards, commentary} so new skills need no new endpoints. 12-round / 4-minute bound; the response write deadline is extended via http.NewResponseController past the 90s server write timeout (required a statusWriter.Unwrap).
• ADR split is the first skill: /api/ai/stories keeps its exact response contract but internally drives the loop; the model calls propose_card once per story it actually finds. The "at most N stories" prompt is gone — the count anchor with it; the cap is now purely server-side.
• Tool scoping: paths that ingest third-party-authored text (the forge-issue ADR variant) run with a read-only tool scope — no update_task, no fetch_link — bounding prompt-injection blast radius. /api/ai/run-skill (user-authored input) gets the full set.
• gpt-5/o-series models get max_tokens renamed to max_completion_tokens at the transport on the runner path, matching the single-shot path's usesMaxCompletionTokens behavior.

Story draft and import stay on the legacy single-shot pipeline; migrating them and deleting that pipeline is #69.

Review findings already addressed

Built via a multi-agent workflow with adversarial review; fixes landed before this PR: the write-deadline no-op through withLogging, side-effect tools reachable from third-party text, partial results discarded on budget errors (cards proposed before an iteration/output limit now return 200 with the failure noted in commentary), and fetch_link inheriting the forge private-address opt-in.

Testing

• go build / go vet / gofmt clean; full suite go test ./...: 1247 tests, 6 packages, 0 failed (also race-clean).
• Go coverage 97.1% (gate 96.4%); per-package all above the 95% floor.
• Runner covered end-to-end against a scripted multi-turn fake OpenAI backend: tool rounds, card cap, partial results, tool scoping, budget-field rename, deadline plumbing (mutation-tested: reverting each fix fails its test).
• Not verified: live gpt-5/o-series upstream (no key in this environment); real-model behavior of the adr-split skill against Ollama Cloud — worth a manual pass before release.

Closes #68.

[socket-security]

Review the following changes in direct dependencies. Learn more about Socket for GitHub.

Diff	Package	Supply Chain Security	Vulnerability	Quality	Maintenance	License
	golang/​github.com/​RandomCodeSpace/​rig@​v0.1.0

View full report

[sonarqubecloud]

Quality Gate passed

Issues
9 New issues
0 Accepted issues

Measures
0 Security Hotspots
97.8% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

[aksOps]

Live validation against Ollama Cloud (qwen3.5:397b — no qwen3.6 exists on this account; closest available flagship):

• POST /api/ai/test: ok:true (tool-calling probe), 1.7s
• POST /api/ai/stories, 2 KiB ADR with 9 deliverables, max=10: 9 stories in 32.0s — content-driven count, no anchoring to the cap; all cards carried prio/effort/tags/checks
• POST /api/ai/run-skill (skill=adr-split, same input): 10 cards + markdown commentary in 31.9s
• POST /api/ai/stories with max=3: exactly 3 stories in 34.2s — server-side cap holds

All four requests HTTP 200, no retries needed. gpt-5/o-series budget-field rename remains unverified against a live upstream.

[aksOps]

Second live validation, gemma4:31b (Ollama Cloud):

• POST /api/ai/test: ok:true, 0.7s
• POST /api/ai/stories, same 9-deliverable ADR, max=10: 9 stories in 7.3s
• POST /api/ai/run-skill (adr-split): 9 cards + commentary in 7.2s
• POST /api/ai/stories, max=3: exactly 3 in 10.0s

All HTTP 200 first try. Both validated backends (qwen3.5:397b, gemma4:31b) independently produced 9 stories from the 9-deliverable document under a cap of 10 — the count now tracks content, not the cap. Gemma omits emoji and writes leaner checklists; all cards still pass validation.