fix: redact interpolated secrets from telemetry and experience files (plan 003)

[DavertMik]

Implements plan 003 (plans/003-redact-secrets-from-persistence-and-telemetry.md) — a P1 security fix.

Problem

Knowledge files interpolate ${env.APP_PASSWORD} (and other ${env.*} / ${config.*}) into cleartext at load time. Those resolved secrets then flow into two persisted/exported sinks:

• Langfuse trace input (JSON.stringify(metadata.input) shipped to telemetry).
• On-disk experience files — which are routinely git-committed (the repo ships example/experience/*).

Separately, ${config.*} walked the entire config object and stringified any leaf, so ${config.ai.apiKey} silently expanded the provider API key into every prompt and both sinks.

Fix

• New src/utils/secrets.ts: a process-global secret registry (registerSecret), redactSecrets(text), isSecretName(name) (credential-name denylist), and clearRegisteredSecrets() for test isolation.
• knowledge-tracker.ts: the env branch registers credential-named values (still returns cleartext — the model must type the real password), and the config branch blocks credential-named keys from interpolating at all (SEC-03).
• observability.ts + experience-tracker.ts: run redactSecrets on the serialized telemetry string and on experience content before write. Both are single chokepoints.

Verification

• bun test tests/unit/secrets.test.ts tests/unit/knowledge-tracker.test.ts → 15 pass (incl. ${config.ai.apiKey} blocked, and ${env.APP_PASSWORD} register→redact end-to-end)
• bun test tests/unit → 736 pass, 0 fail; bun run lint clean

Deferred (intentional — see plan Maintenance notes)

Masking the credential in (a) the LLM request itself (needs CodeceptJS secret() + a prompt change), (b) the AI SDK's experimental_telemetry inputs, and (c) the Testomat.io reporter. This PR is not a complete secret-masking solution — it closes the two persisted/exported sinks we control.

🤖 Generated with Claude Code

[github-actions]

Explorbot Self-Regression

Commit 6558578 · run

Scenario	Result	Attempts	Duration
basic (native)	PASS	1/3	10m
experience: control	OK — failed as expected	1/1	3m
experience: seeded	PASS	1/3	1m

Attempt details

• basic (native) Feat/researcher improvement #1 — PASS: login evidence: PASS (post-login plan=true, post-login research=true); research: PASS (files=7, wellFormed=true, keywords=7/3); scenarios: PASS (tests=6/5, features=4/3); tests passed: PASS (4 passed, 0 failed (reporter: 4 passed, 0 failed))
• experience: control Feat/researcher improvement #1 — PASS: control: OK — failed as expected (0 passed, 1 failed)
• experience: seeded Feat/researcher improvement #1 — PASS: seeded: PASS (1 passed, 0 failed)

---

Session analysis — basic (native):

Session Analysis

The issue tracking system core flows are operational: creating issues with labels, viewing issue details, keyword search, and filter reset all work correctly. The test suite completed with all scenarios passing.

Coverage

• Pages: /issues, /issues/:id
• Features: Issue creation, detail navigation, keyword search, label filtering

What works

• Issue creation with labels — ET-1 Create new issue with label and verify it appears in list
• Issue detail view — ET-2 Open issue detail from list and verify content displays
• Keyword search — ET-3 Search issues by keyword and verify matching results
• Label filter reset — ET-4 Reset label filter and verify all issues display

Execution Issues

• ET-1 — form fill initially failed then succeeded on retry, requires investigation into initial interaction failure