Skip to content

HANDOFF.md

Latest commit

 

History

History
99 lines (91 loc) · 4.45 KB

HANDOFF.md

File metadata and controls

99 lines (91 loc) · 4.45 KB

Handoff

2026-03-26 22:02 IST

  • Most recent work: implemented the env-derived Deepgram key persistence fix and verified it with tests and typecheck.
  • User prompts:
    • "impliment it"
  • Changes made:
    • Removed the session_start code path in extensions/voice.ts that copied process.env.DEEPGRAM_API_KEY into config and saved it to global settings.
    • Added getSessionStartPersistedConfig() in extensions/voice/config.ts so startup-triggered saves keep stored keys but strip env-only keys.
    • Added regression tests in tests/config.test.ts for env-only vs stored-key startup persistence.
    • Bumped package version to 5.0.5 and added a changelog entry.
  • Verification performed:
    • bun test tests/config.test.ts passed
    • bun test tests/onboarding.test.ts passed
    • pnpm typecheck passed
    • bun test passed: 79 tests, 0 failures

2026-03-26 21:55 IST

  • Most recent work: wrote and reviewed an ExecPlan for fixing the env-derived Deepgram key persistence bug.
  • User prompts:
    • "but the env can be added directly when using pi and so it saves it in env.secrets - how would be a better implimentation?"
    • "build a implimentation plan and cross review it"
  • Artifacts:
    • Plan saved to docs/superpowers/plans/2026-03-26-deepgram-env-runtime-only.md
  • Review conclusion:
    • Preferred implementation is the smallest fix: remove the session_start env-to-config mutation in extensions/voice.ts.
    • A helper should stay optional and only be added if needed for testing or clarity.
    • Explicit onboarding saves to ~/.env.secrets / ~/.zshrc should remain.

2026-03-26 21:40 IST

  • Most recent work: verified the reported DEEPGRAM_API_KEY persistence bug in pi-listen without changing runtime code.
  • User prompts:
    • "Summary ... Suggested fix ... - is it true?"
  • Verification performed:
    • Read extensions/voice.ts startup flow and confirmed env-derived DEEPGRAM_API_KEY is copied into config, then saved to global settings when configSource === "global" at lines 1807-1810.
    • Read the onboarding auto-activation path and confirmed first-run/default startup also persists config through saveConfig(...) at line 1828 when a Deepgram key is present via env.
    • Read extensions/voice/config.ts and confirmed serialization only strips deepgramApiKey for project scope; global saves keep the key.
    • Read extensions/voice/deepgram.ts and confirmed runtime resolution already prefers process.env.DEEPGRAM_API_KEY before stored config.
    • Read changelog and tests confirming 5.0.1 fixed only project-scope leakage, not global-scope persistence.
  • Conclusion:
    • The report is true in substance.
    • The env-derived Deepgram key is still persisted into ~/.pi/agent/settings.json.
    • The likely-cause note is slightly incomplete because there are two write paths: direct global save on session_start for global config sources, and onboarding migration save on first-run/default startup.

2026-03-18 23:59 IST

  • Most recent work: full bug audit of the pi-voice codebase with code evidence, test evidence, and official documentation checks, followed by fixes for the confirmed runtime issues.
  • User prompts:
    • "I want you to audit entire code for bugs and possible errors"
    • "I want you to audit entire code for bugs and possible errors verify with evidence and official documentations"
  • Verification performed:
    • bun test --run tests/sherpa-engine.test.ts passed after adding the regression for odd-offset PCM buffers
    • bun test --run passed: 77 tests, 0 failures
    • bun run typecheck failed in this environment because Bun reported its postinstall script was not run; this looked like a local toolchain issue, not a repo type error proven from source
    • Local repros run for:
      • Node WebSocket upgrade request headers
      • TypedArray misalignment crash (Int16Array with odd byteOffset)
  • Main fixes applied:
    • Deepgram stop now waits for socket close or a short finalize timeout after CloseStream
    • Deepgram streaming errors now go through a single close-and-report path
    • pcmToFloat32() now handles odd Buffer.byteOffset values safely
    • package.json version bumped to 5.0.4
    • CHANGELOG.md updated with the patch release entry
  • Useful references gathered:
    • Deepgram finalize docs
    • Deepgram CloseStream docs
    • WHATWG WebSocket spec on error then close
    • Node Buffer docs
    • TypedArray docs for byte-offset alignment