Summary

Ships Stage 5a of agentkeys: the provisioner's deterministic + patterns tier (v0 critical path). agentkeys.provision(service) can now sign up for OpenRouter via a local Playwright script, verify the obtained key against the live API, and store it through the backend. Per the 2026-04-16 CEO review (SELECTIVE EXPANSION), Stage 5b (agentic fallback + audit + script-gen skill) and Stage 6 (npm packaging) are explicitly deferred to v0.1.

• 15/15 PRD user stories PASSED — see .omc/prd.json
• 82 tests passing (67 Rust across 4 crates + 15 TypeScript)
• Architect review: CONDITIONAL_APPROVAL (Opus tier, 4 non-blocking follow-ups routed to Stage 5b, see progress.txt)
• harness/stage-5a-done.sh exits 0 with "STAGE 5a PASSED"
• Rebased onto main clean — crates/agentkeys-core/src/session_store.rs is not part of this PR (no delta from main)

What shipped

Rust (4 crates)

• agentkeys-types::ProvisionEvent + TripwireKind + ProvisionErrorCode — shared schema for Rust↔TS IPC (tagged unions, snake_case)
• agentkeys-provisioner (new crate): Provisioner RAII mutex with PROVISION_IN_PROGRESS sentinel, spawn_and_collect subprocess driver with line-delimited JSON IPC + 120s timeout + malformed-line abort, run_provision orchestrator (verify→store gate), mask_key helper, structured metrics emitter
• agentkeys-mcp::agentkeys.provision tool registered with {service, force?} schema, structured error code mapping
• agentkeys-cli::provision <service> [--force] subcommand with masked-key stdout + CLAUDE.md-format error output

TypeScript workspace (new: provisioner-scripts/)

• src/types.ts — exact mirror of Rust ProvisionEvent (typed boundary, no opaque parsing)
• src/lib/email.ts — Gmail plus-addressing IMAP via imapflow + structured timeout
• src/lib/verify.ts — per-service verification (openrouter → GET /api/v1/models, Bearer auth); returns phantom / endpoint_down / rate_limited distinctions
• src/patterns/signup_email_otp.ts — reusable pattern, zero service-specific strings (grep-enforced)
• src/scrapers/openrouter.ts — composes the pattern + calls verify before emitting Success
• tests/fixtures/openrouter/ — HAR-equivalent hermetic fixtures served via page.route() + route.fulfill()
• tests/scrapers/openrouter.phantom.test.ts — chaos test proving the verification gate catches a sk-or-v1-FAKE decoy key

Harness

• harness/stage-5a-done.sh — 4-gate verifier (cargo tests, npm tests, grep guard against service strings in patterns/, isolated phantom chaos)
• harness/features.json + harness/progress.json updated

Design artifacts (earlier turns this session)

• docs/spec/plans/development-stages.md restructured: Stage 5 → 5a/5b split, Stage 6 deferred, eng-review + design-review decisions locked as implementation notes
• TODOS.md added (Twitter deferred, Phase C scraper target list)
• ~/.claude/skills/agentkeys-record-scraper/SKILL.md (global, not in this PR) — dev-time tool for adding future scrapers

Architectural constraints enforced

• Typed parameters at Rust↔TS boundary; Rust ProvisionEvent enum + TS discriminated union share the exact same JSON shape
• Patterns contain zero service-specific strings (grep -iE "openrouter\|brave\|jina\|groq\|anthropic\|gemini" provisioner-scripts/src/patterns/ returns empty)
• Every scraper success path calls verify() before emitting Success; phantom chaos test proves the gate
• HAR-equivalent fixture-based hermetic tests; zero live-site calls in the test suite
• Pattern extraction regression seam: the same fixture test passed both before AND after extracting signup_email_otp from the inline OpenRouter scraper

Stage 5b follow-ups (architect review, non-blocking)

Logged in progress.txt for Stage 5b:

1. re_verify_existing is a placeholder returning true — real re-verification needed
2. CLI does not stream ProvisionEvent::Progress events to stderr in real time (single-shot today)
3. Phantom chaos emits code: "store_failed" — should be a dedicated ProvisionErrorCode::VerificationFailed variant
4. HAR fixtures are hand-crafted HTML (functionally equivalent, documented) — optional normalization to .har

Optimality suggestions also captured for Stage 5b: streaming orchestrator API, consolidate service-dispatch, NoopBackend extraction, exhaustive event_to_error match.

Scope explicitly NOT in this PR

• Stage 5b: agentic fallback via MCP browser primitives, fallback audit trail, fallback→PR loop, LLM script-generator dev tool usage, 4 additional patterns (OAuth-Google/GitHub, magic-link, password+email-verify)
• Stage 6: npm package, install.sh, README polish
• Stage 8: production hardening
• Additional scrapers beyond OpenRouter — Brave Search, Jina Search, Anthropic, Groq, Gemini come via /agentkeys-record-scraper skill after this merges (per TODOS.md Phase C list). Twitter deferred to TODOs, Instagram dropped.

Test plan

• bash harness/stage-5a-done.sh exits 0
• cargo test -p agentkeys-types -p agentkeys-provisioner -p agentkeys-mcp -p agentkeys-cli passes (67 tests)
• npm install --prefix provisioner-scripts && npm test --prefix provisioner-scripts passes (15 tests)
• grep -iE "openrouter|brave|jina|groq|anthropic|gemini|twitter|instagram" provisioner-scripts/src/patterns/ returns empty
• Phantom chaos test catches a fake-shaped key: cd provisioner-scripts && npx vitest run tests/scrapers/openrouter.phantom.test.ts
• (Manual, requires real Gmail + OpenRouter ToS check) Live provision end-to-end — deferred until ToS clarity per TODOS.md

Reviews already run (2026-04-16)

• /plan-ceo-review — SELECTIVE EXPANSION, 7 proposals, 5 accepted, 1 deferred, 1 rejected; 1 critical gap caught (silent-corrupt-credential → mandatory verification fix) — CEO plan at ~/.gstack/projects/litentry-agentKeys/ceo-plans/2026-04-16-stage-5-hybrid-agentic.md
• /plan-eng-review — 3 P1 decisions locked (IPC schema, concurrency, MCP scope), 5 implementation notes baked into plan, 11 additional tests added to story tables — test plan at ~/.gstack/projects/litentry-agentKeys/*eng-review-test-plan*.md
• /plan-design-review — 4 UX decisions locked (masked-key output, stderr progress, atty TTY detection, duplicate-provision verify-and-report); score 5→9/10
• Ralph architect review — CONDITIONAL_APPROVAL with 4 non-blocking Stage 5b follow-ups

🤖 Generated with Claude Code