- Changed current step from 15 (Performance Test) to 9 (New Task) in _docs/_autodev_state.md, reflecting the transition to Cycle 3. - Updated cycle count from 2 to 3 and modified sub-step details to indicate progress in gathering feature descriptions. - Added new lessons to _docs/LESSONS.md, emphasizing best practices for API key management, dependency handling, and reporting inline fixes during security audits. - Enhanced CI/CD pipeline documentation in _docs/02_document/deployment/ci_cd_pipeline.md to include new gates for vulnerability scans and SBOM emissions, along with dependency overrides for transitive dependencies. - Expanded environment strategy documentation in _docs/02_document/deployment/environment_strategy.md to include the new Google Geocode API key management. Co-authored-by: Cursor <cursoragent@cursor.com>
14 KiB
Retrospective — 2026-05-12 (Phase B Cycle 2)
Mode: cycle-end (autodev existing-code Step 17)
Scope: Phase B, cycle 2 (state.cycle = 2)
Epic: AZ-497 (Self-Hosted Satellite Tiles — SPA Integration) + ad-hoc security tickets AZ-501 / AZ-502 spawned by Step 14
Cycle duration: 2 batches over 1 working day (2026-05-12)
Previous retro: _docs/06_metrics/retro_2026-05-12.md (cycle 1, same calendar day)
Implementation Summary
| Metric | Value | Δ vs cycle 1 |
|---|---|---|
| Total tasks | 4 (AZ-498, AZ-499, AZ-501, AZ-502) | +2 (+100 %) |
| Total batches | 2 (batch 11 = AZ-498 + AZ-499; batch 12 = AZ-501 + AZ-502 inline-fix sub-step under Step 14) | 0 |
| Total complexity points | 11 (AZ-498=5, AZ-499=2, AZ-501≈2, AZ-502≈2) | +1 (+10 %) |
| Avg tasks per batch | 2 | +1 |
| Avg complexity per batch | 5.5 | +0.5 |
| Source files mutated | 12 production + 1 e2e harness + 4 i18n/MSW + 2 scripts + 4 test files + 9 docs | n/a (different shape vs cycle 1's refactor focus) |
Sources: _docs/03_implementation/batch_11_report.md, _docs/03_implementation/batch_12_report.md, _docs/03_implementation/test_run_report_phase_b_cycle2.md, _docs/03_implementation/deploy_planning_sync_cycle2.md.
Quality Metrics
Code Review Results
| Verdict | Count | Percentage | Δ vs cycle 1 |
|---|---|---|---|
| PASS | 0 | 0 % | −2 |
| PASS_WITH_WARNINGS | 1 | 50 % | +1 |
| FAIL | 0 | 0 % | 0 |
| (no formal review — security inline-fix sub-step) | 1 | 50 % | n/a |
Note: batch 12 (AZ-501 + AZ-502) was executed as a Step-14 inline-fix sub-step, not as a Step-10 implement batch, so it did not pass through the implement skill's per-batch self-review path. Static + fast tests covered all 5 ACs implemented in code; the manual-deliverable ACs (AC-6 / AC-7) cannot be verified by tests at all.
Findings by Severity (code review only — security-audit findings tracked separately below)
| Severity | Count | Δ vs cycle 1 |
|---|---|---|
| Critical | 0 | 0 |
| High | 0 | 0 |
| Medium | 0 | 0 |
| Low | 1 (F1 — trim-trailing-slash idiom duplication; pre-existing pattern across 4 call sites in 2 vite roots; consolidation deferred to a future shared-helper extraction task) |
+1 |
Findings by Category (code review)
| Category | Count | Top Files |
|---|---|---|
| Bug | 0 | — |
| Spec-Gap | 0 | — |
| Security | 0 (in code review) | — |
| Performance | 0 | — |
| Maintainability | 1 (Low, pre-existing) | src/features/flights/types.ts, mission-planner/src/services/{Weather,Geocode}Service.ts, src/features/flights/flightPlanUtils.ts |
| Style | 0 | — |
| Scope | 0 | — |
Security-Audit Findings (Step 14, separate from code review)
12 findings total. Inline-fixed this cycle:
| ID | Severity | Status |
|---|---|---|
| F-SAST-1 (Google Geocode key in mission-planner port-source) | HIGH | RESOLVED (AZ-501) |
| F-DEP-1 (Vite ≤ 6.4.1 + PostCSS < 8.5.10 dev-only WebSocket file-read CVEs) | HIGH | RESOLVED (AZ-502) |
F-SAST-2 (unpkg.com CDN ref) |
MEDIUM | DEFERRED (Phase B follow-up) |
F-SAST-3 (STC-SEC2 coverage gap) |
MEDIUM | DEFERRED |
| F-SAST-4 (third-party tile fallbacks) | LOW | DEFERRED |
F-INF-1 (no CI bun audit gate) |
MEDIUM | DEFERRED (tracked in _docs/05_security/infrastructure_review.md) |
| F-INF-2 (missing nginx headers + log redaction) | MEDIUM | DEFERRED |
| F-INF-3 (no Trivy image scan) | MEDIUM | DEFERRED |
| F-INF-4 (no SBOM + cosign signing) | MEDIUM | DEFERRED |
| F-INF-5 (nginx as root, no HEALTHCHECK) | MEDIUM | DEFERRED |
| F-OWASP-1 (security misconfiguration: nginx headers) | MEDIUM | covered by F-INF-2 |
| F-OWASP-2 (vulnerable & outdated components) | MEDIUM | RESOLVED via F-DEP-1 closure (AZ-502) |
Security verdict trajectory: cycle 2 audit overall verdict was FAIL → after AZ-501 + AZ-502 inline fixes, code-level surface returns to PASS_WITH_WARNINGS (Phase B infrastructure follow-ups remain). All 5 deferred F-INF-* items are tracked as concrete next-cycle backlog candidates, not silent gaps.
Structural Metrics
Source: cycle 1 baseline _docs/06_metrics/structure_2026-05-12.md (no new structural snapshot needed — cycle 2 introduced no architecture changes).
| Metric | Cycle 1 close | Cycle 2 close | Δ |
|---|---|---|---|
| Component count | 12 | 12 | 0 |
| Public-API barrels | 11 / 11 (100 %) | 11 / 11 (100 %) | 0 |
| Commit-time static gates | 31 / 31 PASS | 33 / 33 PASS | +2 (STC-SEC1C, STC-SEC1D) |
| Architecture cycles | 0 | 0 | 0 |
| Architecture findings open (baseline F1–F9) | 7 of 9 | 7 of 9 | 0 |
| Newly introduced architecture violations | 0 | 0 | 0 |
| Net architecture delta this cycle | −2 (improvement) | 0 | — |
Wire-contract assertions (endpoints.test.ts) |
36 | 36 | 0 |
| Fast-profile suite | 209 PASS / 13 SKIP / 0 FAIL | 229 PASS / 13 SKIP / 0 FAIL | +20 PASS, 0 SKIP delta |
| Bundle (gzipped initial JS) | not measured | 290 465 B (~14 % of 2 MB budget) | new metric (NFT-PERF-01 baseline) |
Auto-lesson triggers (per skill Step 1)
- Net Architecture delta > 0? No — delta is 0; no
architectureregression lesson required. - Structural metric regression > 20 %? No — every structural metric held or improved.
- Contract coverage % decreased? N/A —
endpoints.test.tscount held at 36; project still uses code-derived contracts. - New finding category emerged? Yes —
security(Step 14 audit fired for the first time this cycle). One of the lessons below captures the rotation-discipline pattern that resulted.
Efficiency
| Metric | Value | Δ vs cycle 1 |
|---|---|---|
| Blocked tasks (cycle-internal) | 0 | 0 |
| Tasks pending external user action | 2 (AZ-499 AC-7 OWM revocation, AZ-501 AC-6 Google revocation) | +2 (new pattern) |
| Cross-workspace gates outstanding | 1 (AZ-498 deploy via satellite-provider cookie-auth) | +1 (new pattern) |
| Tasks requiring fixes after review | 0 | 0 |
| Batch with most findings | batch 11 (1 Low pre-existing) | n/a |
| Auto-fix loops invoked | 0 | −1 |
| Stuck-agent incidents | 0 | 0 |
Blocker Analysis
| Blocker Type | Count | Prevention |
|---|---|---|
| Manual third-party-console action (key revocation) | 2 | Folded into the new "external-secret" task template (Improvement Action #2 below) |
| Cross-workspace ticket dependency (deploy gate) | 1 | Surface during Step 9 (New Task) when ticket scope crosses workspace boundaries; capture in the task spec's Dependencies field as it was for AZ-498 |
User-decision points (cycle 2 only)
- Step 14 outcome (HIGH findings): user chose A (fix both inline) — produced AZ-501 + AZ-502.
- Step 15 perf: user chose A (run perf tests) — confirmed bundle stays under budget.
- Commit decision: user chose B (commit + push to remote
dev) —f7dd6c9pushed. - Step 16 deploy gate: user skipped the structured choice; agent defaulted to planning-only sync (option B in the absence of an answer) and recorded the prod cutover + key revocations as leftovers. Rationale: the unanswered options A (full deploy) required external state I could not verify, and option C (skip entirely) would have lost the planning information.
Trend Comparison
| Trend | Cycle 1 | Cycle 2 | Direction |
|---|---|---|---|
| Code review pass rate | 100 % | 50 % (1 PASS_WITH_WARNINGS, 1 no-formal-review sub-step) | ⬇ but explainable: PWW finding was pre-existing Low, not introduced this cycle |
| Test count | +46 (cumulative this cycle) | +20 (this cycle on top of cycle 1) | continued positive growth |
| Static gate count | +2 | +2 | continued positive growth (now both axes: arch + security literal-scan) |
| Architecture findings open | 7 (−2) | 7 (0) | held; cycle 2 was config/wire + security, no architecture surface touched |
| Pending USER actions at cycle close | 0 | 2 (revocations) + 1 (cross-workspace gate) | ⬆ — first cycle to exit with non-zero user-action backlog; visible in leftovers |
The cycle 2 user-action backlog is a structural side-effect of running Step 14 (Security Audit) for the first time, not a process regression. The Phase A baseline never scanned for committed secrets; cycle 2's audit surfaced two such secrets that could only be neutralized via vendor-console action. Both are tracked in _docs/_process_leftovers/2026-05-12_az-498-deploy-and-key-revocations.md with full replay procedures.
Top 3 Improvement Actions
-
Run Step 14 (Security Audit) earlier in the cycle, ideally as a pre-flight to Step 9 (New Task). This cycle's audit caught two HIGH findings (Google Geocode key + Vite CVEs) after the implement work was complete, forcing the inline-fix detour and producing AZ-501 / AZ-502 mid-cycle. Running a lightweight static-only audit pre-Step-9 (read
mission-planner/src/config.ts,mission-planner/src/services/, top-level deps) would have surfaced the Google key during AZ-499's planning — bothmission-planner/keys could have been externalized in the same batch as AZ-499.- Impact: high — would have collapsed AZ-499 + AZ-501 into a single batch with a single rotation discipline; would have caught F-DEP-1 before AZ-498 implementation began (cleaner branch state).
- Effort: low — add a
pre-cyclemode to.cursor/skills/security/SKILL.mdthat runs Phase 1 (deps) + Phase 2 (SAST) only, callable from Step 9 of the existing-code flow.
-
Standardize an "external-secret externalization" task template. AZ-499 and AZ-501 are mechanically identical: extract to service module → env var via
import.meta.env.VITE_*→ fail-soft return → add literal-scan static gate (STC-SEC1x) → document in.env.examplewith<your-...>placeholder → leave the actual revocation as a manual deliverable AC. The third such task (whichever comes next) should copy a checklist, not re-derive the pattern.- Impact: medium-high — directly addresses the cycle-2 user-action backlog as a structural pattern; the next external-secret task lands in a single PR with all 6 steps already scoped.
- Effort: low — add
_docs/02_tasks/_templates/external_secret_externalization.md(new) and reference it from.cursor/skills/new-task/SKILL.md's "Task Type Detection" section.
-
Enforce
bun audit --severity highin CI (close F-INF-1). F-DEP-1 (Vite/PostCSS CVEs) was found by manualbun auditinvocation during the audit. A CI gate would have caught it within hours of the advisory being published, instead of waiting for the next manual audit cycle. The fix is small — one Woodpecker step before the build stage — and the AZ-502package.jsonoverrides already make the gate green today.- Impact: medium — closes a known coverage gap before the next dependency CVE lands; pairs with action #1 (security earlier in cycle) to push security from "audit" to "continuous gate".
- Effort: low — single step addition to
.woodpecker/build-arm.yml.
Suggested Rule / Skill Updates
| File | Change | Rationale |
|---|---|---|
.cursor/skills/security/SKILL.md |
Add a pre-cycle invocation mode that runs Phase 1 (deps) + Phase 2 (SAST) only, with a 5-minute time budget. Wire it into .cursor/skills/autodev/flows/existing-code.md as an optional pre-Step-9 gate. |
§Top 3 Improvement Action #1. |
_docs/02_tasks/_templates/external_secret_externalization.md |
NEW file. Template with the 6-step checklist (extract to service module → env var → fail-soft → STC-SECx literal-scan → .env.example placeholder → manual revocation AC). Include AZ-499 and AZ-501 as canonical examples. |
§Top 3 Improvement Action #2. |
.cursor/skills/new-task/SKILL.md (Task Type Detection) |
Add an "external-secret-externalization" trigger phrase set ("hardcoded API key", "rotate credential", "externalize secret") that suggests the new template. | §Top 3 Improvement Action #2 enablement. |
.woodpecker/build-arm.yml |
Add a bun audit --severity high step before the build stage (closes F-INF-1). |
§Top 3 Improvement Action #3 + audit infrastructure_review.md F-INF-1. |
_docs/LESSONS.md (top) |
Append the 3 lessons in §LESSONS Append below; trim to ≤ 15 entries. | Skill Step 4. |
Notes — Step 16 outcome
Step 16 (Deploy) ran in planning-only mode because:
- The user skipped the structured deploy-gate choice; the agent defaulted to option B (plan only) since option A required unverifiable cross-workspace state and option C would have lost the planning information.
- The actual prod cutover for AZ-498 + the two key revocations are tracked as leftovers — see
_docs/_process_leftovers/2026-05-12_az-498-deploy-and-key-revocations.md(3 entries, each with a full replay procedure). _docs/02_document/deployment/{environment_strategy,ci_cd_pipeline}.mdwere updated to reflect cycle 2 changes (new env var + override block) so the next cycle's Step 16 starts from accurate planning artifacts.
LESSONS Append (top 3, single-sentence, tagged)
- [process] When externalizing a committed API key, always follow the 4-step rotation discipline: (a) extract to env-var via a service module so unit tests can stub it, (b) add a literal-scan static gate (STC-SECx) against the rotated value as defense-in-depth, (c) document in
.env.exampleusing the established<your-...>placeholder convention, (d) leave the actual key revocation as a manual deliverable AC with evidence-attachment requirement — never assume the static gate alone neutralizes the leaked credential. - [dependencies] When
bun auditreports advisories on a transitive dep that directbun update <dep>does not clear (because nested copies persist under sibling tools, e.g.vitest/node_modules/<dep>), usepackage.json"overrides"to floor the resolution AND clean reinstall (rm -rf node_modules bun.lock && bun install) — a direct update alone cannot displace nested copies, and Bun honors the npm-compatibleoverridesfield exactly as npm does. - [tooling] When the autodev orchestrator delegates to a sub-skill that ends in a HIGH-severity blocking gate (e.g. security audit FAIL → user picks "fix inline"), capture the inline-fix sub-step results as a separate batch report (
batch_NN_report.md) — not as an extension of the prior batch — so the cycle metrics correctly attribute findings, ACs, and complexity to the work boundary that produced them.