Hiro development journal

Bounded Moltbook and Reddit idea sources prepared without disturbing Stage 6C

2026.08.11 at 08:31:12 America/Los_Angeles Implementation tested and isolated; live activation deferred at the Stage 6C revision boundary Machine-readable JSON

Executive summary

A bounded external-idea lane was implemented for Hiro's continuous improvement loop with exact adapters for Moltbook and Reddit.

Moltbook's public read-only posts endpoint is supported immediately. A live preview made one permitted GET, reviewed 20 posts, and produced three local experiment questions.

Reddit support uses only its official OAuth API and read scope. It remains fail-closed until Reddit API approval, an approved application identity, credentials, and an identifying user agent are configured; the preview made no Reddit request.

External text is treated as untrusted evidence and reduced deterministically to curiosity hypotheses. It cannot directly create candidates, modify code, grant authority, or enter Stage 6.

The tested change is committed on an isolated branch. Hiro's live checkout remains clean on the exact revision bound to the running Stage 6C campaign, which still has zero recorded shadow decisions.

Work completed

External-source policy and adapters

Implemented in isolation

Added a versioned policy with exact Moltbook and Reddit endpoints, small response and time budgets, a maximum of three ideas per cycle, and a local-corroboration requirement.
Moltbook permits one unauthenticated GET to its exact public posts endpoint with the existing pinned-address, public-network-only observation client.
Reddit permits only the official OAuth token endpoint and OAuth content origin, a fixed subreddit allowlist, and the read scope.
Reddit tokens, credentials, raw response bodies, authors, and account identifiers are not retained in idea packets.
Neither adapter implements posting, commenting, voting, following, messaging, uploads, arbitrary browsing, or another state-changing action.

Untrusted-content reduction

Implemented and tested

External content is sanitized, secret-like material is redacted, and obvious instruction-injection patterns are rejected.
A deterministic reducer maps relevant terms into bounded themes such as evaluation reliability, provenance, memory, latency, rollback, routing, and observability.
The retained output is a question, rationale, local experiment, source URL, and content hash rather than an instruction copied from the external post.
Each accepted idea remains a curiosity hypothesis. Local evidence and a complete ImprovementSpec are required before isolated candidate construction can begin.
Deduplication prevents the same content identity from being accepted repeatedly.

Continuous-loop integration

Implemented in isolated revision

External hypotheses are collected during the existing proposal-only curiosity phase, ahead of locally generated curiosity items.
The active-loop state records the number of external sources checked, items reviewed, and ideas created.
External-source failures are isolated from the rest of the improvement cycle and fail closed.
The change does not revive automatic Daylab, Nightlab, or the retired internet-observation probation.

Stage 6C revision-boundary protection

Live checkout preserved

The running Stage 6C validation campaign is bound to the current live revision and still records zero of 30 required shadow decisions.
Advancing and restarting the live checkout on the new external-source revision would make that exact campaign revision binding false.
The complete tested change was therefore committed in a separate worktree and branch while the live checkout was restored to a clean state on its campaign-bound revision.
No Stage 6C state, ledger, runtime fragment, service, or scheduler was changed during this preparation.

Decisions and reasoning

Interpret the user's request as authorization for read-only idea collection, not for social actions or account impersonation.
Activate Moltbook through its exact public read endpoint but require Reddit's own approval and OAuth configuration before any Reddit network request.
Prevent external content from directly authorizing or specifying code changes; require local corroboration and the existing evaluation pipeline.
Preserve the exact Stage 6C revision binding by keeping the tested change isolated until an explicit campaign transition decision is made.
Do not restart Hiro from uncommitted code or silently rewrite the durable Stage 6C campaign state.

Validation and evidence

Check	Status	Result
Focused external-source and observation suites	passed	64 focused tests passed, covering policy bounds, deterministic reduction, hostile instruction rejection, deduplication, Reddit fail-closed behavior, continuous-loop integration, and the existing observation safety suite.
Full Hiro repository suite	passed	The final full repository suite passed all 471 tests in 157.65 seconds.
Moltbook live preview	passed	One bounded GET reviewed 20 public posts and produced three hypotheses concerning freshness, rollback hysteresis, and disposable runtimes. Preview mode did not consume dedup state.
Reddit live access	correctly blocked	The adapter reported configuration_required because Reddit API approval was not confirmed. It made no Reddit content request.
Live Hiro and Stage 6C state	preserved	The live checkout remains clean on the campaign-bound revision. The Stage 6C campaign remains running in shadow-decisions phase with zero counters and no active runtime fragment.
Journal test suite	passed	The required timestamped-entry journal tests completed successfully.
Journal production build	passed	The required journal generation, output validation, TypeScript compilation, and production build completed successfully for 101 entries.

Current state

The external-source implementation is frozen in isolated revision 2d86ad6.
Moltbook retrieval and idea reduction have passed a real read-only preview.
Reddit code is ready but cannot legally or technically retrieve content until the private runtime is configured with approved read-only API access.
Hiro's live process has not been restarted with the new code.
The existing Stage 6C campaign remains unchanged and bound to the prior live revision.

Next steps

Choose whether to preserve the current Stage 6C campaign until completion or supersede its zero-progress state and restart validation on the external-source revision.
If immediate activation is chosen, freeze a new exact revision binding, rerun the required validation, restart through the hidden launcher, and verify the first persisted external-idea cycle.
Register and obtain approval for a non-commercial Reddit Data API application before configuring the private read-only OAuth values.
Add a bounded local-corroboration stage that can promote a promising external curiosity item into a complete ImprovementSpec only when Hiro can reproduce a relevant local gap.
Expose source availability and accepted external themes on the benchmark page without showing raw external content or credentials.