Authenticated Hermes Agent backend with private continuous Timmy chat #48

New Issue

timmy · 2026-08-20T15:23:48Z

timmy commented

2026-08-20 15:23:48 +00:00

Parent epic: #6

Depends on the simplified Timmy shell and must preserve #5 safety/privacy boundaries.

Outcome

Replace canned prompt buttons with a smart conversational Timmy backed by the full server-side Hermes Agent while preventing an unauthenticated browser from spending privileged tool authority.

Acceptance criteria

Hermes credentials and real Hermes session IDs never reach browser JavaScript
Agent is disabled by default and requires explicit server configuration plus authenticated browser access
Browser sessions map to opaque server-owned Hermes sessions with continuity
Same-origin, body/time/concurrency/rate limits, cancellation, sanitized errors, and expiry fail closed
Medical system brief prohibits diagnosis, symptom invention, autonomous treatment, and suppression of deterministic escalation
Smart chat supports free text and ledger-aware context, with graceful local fallback
Adversarial tests cover unauthorized use, session theft, prompt attempts to change tool policy, oversized input, and timeout
Release demo shows real chat continuity without exposing sensitive data

Required evidence

RED/GREEN tests, exact CLI/runtime receipt, API probes, mobile visual proof, and full suite.

Parent epic: #6 Depends on the simplified Timmy shell and must preserve #5 safety/privacy boundaries. ## Outcome Replace canned prompt buttons with a smart conversational Timmy backed by the full server-side Hermes Agent while preventing an unauthenticated browser from spending privileged tool authority. ## Acceptance criteria - [ ] Hermes credentials and real Hermes session IDs never reach browser JavaScript - [ ] Agent is disabled by default and requires explicit server configuration plus authenticated browser access - [ ] Browser sessions map to opaque server-owned Hermes sessions with continuity - [ ] Same-origin, body/time/concurrency/rate limits, cancellation, sanitized errors, and expiry fail closed - [ ] Medical system brief prohibits diagnosis, symptom invention, autonomous treatment, and suppression of deterministic escalation - [ ] Smart chat supports free text and ledger-aware context, with graceful local fallback - [ ] Adversarial tests cover unauthorized use, session theft, prompt attempts to change tool policy, oversized input, and timeout - [ ] Release demo shows real chat continuity without exposing sensitive data ## Required evidence RED/GREEN tests, exact CLI/runtime receipt, API probes, mobile visual proof, and full suite.

timmy added this to the M3 — Private Beta & Longitudinal Value milestone 2026-08-20 15:23:48 +00:00

timmy added the

labels 2026-08-20 15:23:48 +00:00

timmy self-assigned this 2026-08-20 15:23:48 +00:00

timmy referenced a pull request that will close this issue

2026-08-20 16:21:31 +00:00

Sleek Timmy mobile shell and secure Hermes Agent chat #49

timmy closed this issue

2026-08-20 16:25:48 +00:00

timmy reopened this issue

2026-08-20 16:35:36 +00:00

timmy commented

2026-08-20 16:35:36 +00:00

Independent post-merge review found that deterministic urgent-symptom interception existed only in browser code, omitted common phrases, and did not inspect ledger flags at the authoritative server boundary. Hotfix in progress with red/green regression tests at domain, service, HTTP, and browser layers.

timmy referenced a pull request that will close this issue

2026-08-20 16:36:18 +00:00

Hotfix: enforce deterministic urgent-symptom override at the server boundary #50

timmy closed this issue

2026-08-20 16:48:03 +00:00

timmy reopened this issue

2026-08-20 16:56:21 +00:00

timmy commented

2026-08-20 16:56:21 +00:00

Follow-up hostile review found a remaining common-phrase bypass: I threw up was not recognized. Release daily-2026-08-20.2 is superseded. A second red/green hotfix is in progress with domain, authoritative service, and browser coverage.

Follow-up hostile review found a remaining common-phrase bypass: `I threw up` was not recognized. Release daily-2026-08-20.2 is superseded. A second red/green hotfix is in progress with domain, authoritative service, and browser coverage.

timmy referenced a pull request that will close this issue

2026-08-20 16:57:27 +00:00

Hotfix: recognize past-tense vomiting in deterministic chat safety #51

timmy closed this issue

2026-08-20 17:17:34 +00:00

Sign in to join this conversation.