Hiro development journal

Routing ideas become specific, aggregated, and evidence-ranked

2026.08.15 at 13:11:47 America/Los_Angeles Implemented, migrated, validated, and live Machine-readable JSON

Executive summary

Hiro's queue had admitted many external source leads as separate records even when all of them reduced to the same generic tool-routing sentence. Because priority used only the shared theme and one-source support count, those records also received identical totals.

External routing evidence now maps to one of five finite, code-owned approaches: fallback recovery, resolver matching, workflow orchestration, planning handoff, or capability matching. Raw external prose remains excluded from prompts and queue summaries.

The actionable queue now aggregates source leads by approach, reports supporting sources and an external evidence signal, and calculates priority from both evidence strength and approach-specific benefit, cost, risk, and complexity.

Seven queued generic routing records were moved to Superseded with append-only reasons. Fresh discovery reviewed seventy-two external items and produced an approach-specific capability-matching lead; the next admission created five aggregated approach records across the available mechanisms.

The live Waiting view contains the specific capability-and-authority routing proposal and no copy of the former generic routing sentence.

Work completed

Root-cause correction

Completed

Traced the repetition to the prompt-safety reducer, which correctly discarded raw source prose but collapsed every tool, router, workflow, resolver, planning, and orchestration signal into one tool_routing mechanism.
Confirmed that the continuous backlog admitted unclustered reservoir leads and that the priority function ignored each lead's external rank score, producing separate but textually and numerically identical cards.
Preserved the security boundary: external text is still never forwarded to the improvement model, persisted in hypotheses, or treated as instructions.

Code-owned routing approaches

Completed

Added five allowlisted routing approaches with distinct safe proposals, intended behaviors, and metrics.
Fallback recovery tests preferred-route failure and bounded fallback; resolver matching tests registry, schema, and capability matches; workflow orchestration tests ordering and dependencies; planning handoff tests the reasoning-to-tool boundary; capability matching tests least-authority tool selection.
Approach identity is included in the safe lineage key so genuinely different routing mechanisms do not collapse together, while multiple sources supporting the same approach do.

Aggregated queue admission and migration

Completed

Changed automatic admission from one record per source lead to one record per safe approach lineage. Supporting source names and opaque references are combined, and the strongest external signal receives a bounded corroboration increment.
Manual Make next admission now resolves to the same aggregate approach record instead of bypassing deduplication with an individual source lead.
Added an append-only migration that supersedes queued external routing ideas lacking a concrete approach. It does not rewrite terminal evidence or forcibly alter an already active transaction.
The live migration superseded seven generic waiting records and increased the Superseded count from ten to seventeen.

Transparent differentiated priority

Completed

Added external evidence strength as a bounded priority factor that changes confidence and reach.
Added approach-specific capability adjustments and cost, risk, and complexity values, preventing different routing approaches with equal evidence from receiving unexplained identical totals.
Queue cards now show the approach label, supporting-source count and names, external evidence signal, and a note that priority uses evidence strength and corroboration.
After migration, the capability-matching routing record scored 15.63 from a 4.51 external signal, while other admitted approaches scored 28.58, 13.47, 12.27, and 6.06 according to their evidence and work profile.

Decisions and reasoning

Do not preserve external prose merely to make cards look different; derive actionable distinctions through a finite code-owned taxonomy that maintains the prompt-injection boundary.
Aggregate corroborating sources into one actionable hypothesis instead of treating every feed item as independent implementation work.
Reject generic routing evidence as insufficiently actionable rather than letting it refill the queue.
Use external rank only as a bounded evidence factor, not as an instruction or direct authority signal.
Allow an already active generic candidate to conclude through its existing transaction instead of mutating or deleting it mid-run; remove generic duplicates from Waiting immediately.
Develop in an isolated Git worktree while the promotion loop was active, then integrate one validated commit into the clean live branch.

Validation and evidence

Check	Status	Result
Focused queue, source, dashboard, and scoring tests	passed	Fifty-one focused tests passed for approach classification, safe clustering, source aggregation, migration, evidence-based ranking, approach-specific ranking, dashboard output, and existing queue behavior.
Authoritative full Hiro suite	passed	All 583 tests passed in 180.33 seconds in Hiro's real environment with the pinned runtime and frozen model-benchmark bytes.
Live restart	passed	Hiro restarted successfully on Qwen3.8, both API tiers and the Observatory returned, and the circuit breaker remained closed at zero infrastructure failures.
Generic queue migration	passed	Seven queued external tool-routing records without an approach were moved to Superseded with an append-only concrete-approach reason.
Fresh external discovery	passed	The post-restart cycle checked seven sources, reviewed seventy-two items, retained one new approach-specific lead, and completed without external action authority.
Live actionable queue rendering	passed	The Waiting filter displayed capability matching, the specific capability-and-authority proposal, evidence strength, and differentiated priority; the former generic routing sentence was absent and the browser console had zero errors.

Current state

The actionable queue is no longer filled with duplicate generic routing cards.
One capability-matching routing approach is currently available from fresh evidence; additional routing approaches will appear only when source evidence matches their finite definitions.
Five newly admitted aggregate records expose distinct summaries, supporting sources, evidence signals, and priority totals.
Historical generic records remain preserved under Superseded rather than being deleted.
The repository is clean at commit 3c572ac and the continuous-improvement circuit breaker remains closed.

Next steps

Observe future discovery cycles to confirm that resolver, fallback, orchestration, and handoff approaches appear only with matching evidence and aggregate corroborating sources correctly.
Let the already active legacy candidate finish through the normal evaluator, then retain or reject it without special treatment.
Use the new Waiting filter to review actionable approaches separately from preserved Superseded history.