Explore V2: Single-Conversation Explorer With a Schema-Bound Finish

Written 2026-06-09 after reviewing the candidate-selection implementation in src/pro/main/ipc/handlers/local_agent/tools/explore_code_subagent.ts on the explore-code-subagent branch. Builds on GENERALIZABLE_EXPLORER_AGENT.md and LONG_TERM_EXPLORE.md; supersedes their implementation sequencing where they conflict.

Summary

The current implementation got the core invariant right: the value model selects observed candidate IDs and can never type a file path or line range into the final report. That fixed fabricated references.

But the implementation around that invariant has three structural problems:

Laundered overfitting. Benchmark-specific path literals were deleted, but the behavior came back as "generic" English word lists, a web-app role taxonomy, and intent regexes that are visibly derived from the benchmark repos (scene, channel, canvas-style target resolution, /onboarding/ as a support path). These will misfire on mobile apps, backend services, CLIs, and non-React conventions.
Split-brain orchestration. The explorer model builds full understanding across up to 8 tool steps, is told to throw it away (respond only: done), and then a separate selection call re-derives judgment from a lossy 180-char-evidence candidate packet. With the targeted and gap follow-up passes, one explore_code call can cost ~6 sequential LLM invocations, each rebuilding context the first one already had.
A report that repeats itself and begs. The deterministic report renders the same facts up to four times (answer draft, JSON summary, findings, causal chain, flow, recommended action) and embeds 400-600 chars of per-call "do not call grep again" imperatives in uncached tokens — policy that already lives in the cached tool description and system prompt.

V2 collapses exploration and selection into one conversation that must end with a structured submit_report tool call, demotes deterministic code from judge to validator, makes task intent a tool argument supplied by the main model, and cuts the report to a single representation at roughly a third of the current budget.

The fabricated-fact problem also moves into scope: the selection model currently authors causal "facts" from candidate metadata it never read. V2 grounds facts by requiring them to quote observed evidence, and earns main-model trust with short verbatim quotes instead of per-report prohibitions.

Goals

One model invocation chain, not six. Exploration, judgment, and gap-filling happen in a single sub-agent conversation. Follow-up is a continuation, not a re-selection pass.
Generalizability by deletion. No English morphology tables, action-word lists, web-app role enums scored by regex, or query-intent regexes in production scoring code.
Dense, trusted main context. One-representation report at ~2,500 chars with verifiable evidence quotes. Policy text lives in cached context, never in the per-call report.
Deterministic code validates; it does not judge. It can drop, clamp, dedupe, and lower confidence. It never upgrades an action, pads a file list, or rewrites the model's selection.
Keep the core invariant. The model selects candidate IDs only. Unobserved references stay unrepresentable.

Non-Goals

A semantic index or new retrieval channels (Phase 3/5 of the prior plans still apply later).
Letting the model write raw path:line references.
Optimizing sub-agent token spend at the expense of main-model context density. The sub-agent is cheap; the main model is not.
Preserving current benchmark scores through the refactor. A dip when word lists are deleted is signal, not regression.

Current State (what to keep, what to delete)

Keep

Candidate ID invariant and resolveCandidateIds-style resolution.
Candidate extraction at the tool boundary (candidatesFrom*Result), dedupe, overlap merging, range clamping (clampRangeForReport), and ranked-candidate ID assignment.
Chat-scoped report cache with file-stat invalidation (explore_code.ts).
Benchmark recorder instrumentation and the new --arms / --resume-from harness plumbing.
The grep literal-inference fallback (grep.ts) — genuinely generic.
Generic test/support/generated/docs path classification, minus product-shaped entries (/onboarding/, /tours/).

Delete

selectExploreCandidates and the standalone selection prompt/parse path (parseExploreSelection, extractJsonObject, and friends) — replaced by a forced tool call.
runTargetedEvidencePass, runModelAuthoredGapEvidencePass, runMinimumEvidencePass, and their query-synthesis helpers (buildMinimumEvidenceQueries, buildTargetedIdentifierGrepQueries, extractTargetIdentifiers) — replaced by conversation continuation.
inferTaskIntent and every intent-conditional branch — replaced by an intent tool argument.
normalizeRecommendedAction — replaced by validation-only rules (see below).
supplementPrimaryCandidates and MIN_ANSWER_PRIMARY_FILES padding.
normalizeSearchTerm morphology table, isActionSearchTerm, isLowSignalSearchTerm, the scene/scenes blocklist entries, and hasTargetResolutionEvidence.
getEvidenceRoles / getStrongEvidenceRoles / hasStrongApiEvidence regex haystacks — role judgment moves to the model with an open vocabulary.
The target evidence role (canvas-app residue).
Answer draft, separate findings list, and duplicated causal-chain prose in the report.
Per-report imperative policy text.
DYAD_CODE_EXPLORER_REPORT_MODE env-var control of production behavior — benchmark arms get explicit plumbing.

Known bugs to fix in passing

Dead ternary in resolveSelection (both branches return fallbackReadTargets).
Duplicate routes? check in normalizeSearchTerm (moot once deleted).
skip_explore_result parses but silently renders as targeted_gap_search — make it a real rendered outcome ("explorer found nothing relevant; proceed without it") or drop it from the schema.
Per-observation char cap exists but total raw-observation budget is unbounded across steps.
estimatedTokens = rangeWidth * 8 overestimates ~2x; use ~4 tokens/line or measure.

Architecture

text

user query + intent (from main model)
  -> sub-agent conversation (single streamText call, tool loop)
       tools: explore_code(raw), grep, read_file, list_files, submit_report
       every tool result is annotated inline with candidate IDs: [c12]
       host accumulates typed candidates at the tool boundary (unchanged)
  -> conversation MUST end by calling submit_report(selection)
       selection schema = candidate IDs + roles + facts + action + confidence + gaps
  -> deterministic validator
       drop unknown IDs / unobserved ranges; clamp budgets; verify fact quotes;
       lower confidence; never upgrade or rewrite
  -> if validator finds a critical gap AND step budget remains:
       append one user message naming the gap; continue the SAME conversation
  -> deterministic renderer emits the final compact report
  -> main model answers, reads listed ranges, or runs the listed bounded searches

1. Candidate IDs injected inline

Today candidates get IDs only after exploration ends, so the explorer model cannot reference them. V2 assigns IDs as observations arrive and annotates the tool result text the model sees:

compiler windows: #### src/store/channels.ts [c7] - switchChannel (function:42)
grep clusters: one ID per rendered cluster;
read_file: one ID for the read range;
list_files: IDs only for paths, marked path-only.

The registry stays host-side and typed (unchanged from today). The inline annotation is purely so the model can select IDs it has actually seen, in the same conversation, with the full evidence in context. ID assignment must be stable across the conversation (monotonic counter, dedupe maps to the first ID).

2. `submit_report` as a forced tool call

Replace the prose-JSON selection pass with a submit_report tool whose input schema is the selection. Enforce completion: if the model stops without calling it, send one nudge message; if it still fails, fall back to deterministic-only selection (current fallback behavior). Schema validation happens at the tool-call layer, so extractJsonObject regex parsing and the selection_invalid path disappear.

interface ExploreSelectionV2 {
  primaryCandidateIds: CandidateId[]; // 1-5, no padding
  readTargets: Array<{
    candidateId: CandidateId;
    purpose: string; // tied to the caller's intent
    required: boolean;
  }>;
  flow: Array<{
    candidateId: CandidateId;
    role: string; // OPEN vocabulary; suggested list in prompt only
    fact: string; // must contain a quote from observed evidence (validated)
    quote: string; // <=2 lines, verbatim from an observed window/cluster
  }>;
  missingCoverage: string[]; // specific, <=3
  recommendedPrimaryAction:
    | "answer_from_report"
    | "read_targets"
    | "targeted_gap_search"
    | "skip_explore_result";
  searchTargets?: string[]; // only for targeted_gap_search; bounded terms+scopes
  confidence: "high" | "medium" | "low";
}

Notes:

role is a free string with a suggested vocabulary (entry, ui, handler, state, data/api, persistence, render/output, type, test) in the prompt. No closed enum, no regex scoring of roles. A mobile app can say "gesture recognizer"; a CLI can say "command dispatch".
quote is the trust mechanism: <=2 verbatim lines per flow link (~30 tokens) that the validator string-matches against observed evidence. A verifiable quote does more to stop main-model re-reading than any amount of prohibition text.
flow replaces findings + causalChain + flowSummary. One list, ordered, is the explanation.

3. Intent comes from the caller

Add to exploreCodeSchema:

intent: z.enum(["explain", "locate", "edit", "debug"]).describe(
  "What the result will be used for. explain/locate: answer or point at code. " +
    "edit/debug: exact ranges will be read before changing code.",
);

The main model knows why it is calling the tool; inferring intent from English query regexes is strictly worse. Intent flows into the sub-agent prompt and into validation thresholds (e.g. answer_from_report is only legal for explain/locate). inferTaskIntent is deleted. Cache key must include intent.

4. Adaptive continuation instead of follow-up passes

After submit_report, the validator checks the selection. If there is a critical, specific gap (a flow link whose quote failed validation, a missing role the model itself named, zero ranged candidates for an edit intent) and the budget allows, the host appends one user message to the same conversation:

text

Your report cited a missing link: "<gap>". You have N tool steps remaining.
Find observed evidence for it, then call submit_report again.

At most 2 continuation rounds. Budgets: ~12 total tool steps, a total raw-observation cap (~60k chars), and wall-time. The model keeps everything it already learned; no packet rebuild, no re-selection call, no separate gap-pass system prompt. Stop reasons are recorded for the benchmark.

5. Validator (the only deterministic policy)

Allowed operations, in order:

drop unknown candidate IDs and read targets without observed ranges;
drop flow links whose quote does not appear (whitespace-normalized substring match) in that candidate's observed evidence; record fact_unverified for each;
clamp counts (<=5 primary, <=8 read targets, <=3 missingCoverage) and ranges (clampRangeForReport);
dedupe and overlap-merge (existing logic);
downgrade only:
- answer_from_report with intent edit/debug -> read_targets (or targeted_gap_search if no ranged targets survive);
- read_targets with zero surviving targets -> targeted_gap_search;
- confidence high with any dropped link or non-empty missingCoverage -> medium;
- confidence medium with zero surviving flow links -> low.

Forbidden operations: upgrading an action, padding primary files, synthesizing search queries, rewriting facts, reordering the model's flow. If validation guts the selection (no primary files survive), fall back to the deterministic low-confidence report.

6. Report format: one representation

text

## explore_code report
Query: "..." | Intent: explain | Confidence: high | Action: answer_from_report

Flow:
1. src/routes/app.tsx:18-44 (entry) - Route mounts <ChannelSidebar/>.
   > <Route path="/channels/:id" element={<ChannelSidebar/>} />
2. src/components/sidebar.tsx:120-163 (handler) - Click calls switchChannel(id).
   > onClick={() => switchChannel(channel.id)}
3. src/store/channels.ts:42-58 (state) - switchChannel dispatches setCurrentChannel.
   > dispatch(setCurrentChannel(id))

Missing: none
Read targets (only if editing): src/store/channels.ts:42-58 - edit the dispatch payload.

```json
{ ...compact machine block: paths/ranges/action/confidence only... }


Rules:

- Each path appears exactly once outside the JSON block.
- The JSON block carries only what machines need (cache invalidation in `explore_code.ts`
  parses it; keep that contract but shrink it — paths, ranges, action, confidence).
- Budget: `MAX_REPORT_CHARS = 2_500` (down from 8,000).
- Zero imperative policy text. "Follow recommendedPrimaryAction", "don't re-explore after a
  high/medium report" live only in the tool description and `local_agent_prompt.ts`, which are
  cached per session. Update both to describe the V2 format and remove references to sections
  that no longer exist (answer draft, findings).
- Quotes are <=2 lines each and are the *only* source text allowed in the report. The validator
  rejects anything longer.

### 7. Ranking stays, word lists go

Until worker-side graph ranking (prior plan Phase 3) lands, keep the existing
`buildCandidate` scoring but reduce it to structural features only:

- source weight (compiler > read_file > grep > list_files);
- exact query-identifier match against path basename / symbol names / evidence — using the
  raw query tokens split on non-alphanumerics and camelCase, **no morphology table, no
  action/noun lists**;
- range tightness and estimated token cost (fixed: ~4 tokens/line);
- generic test/support/generated/docs penalty.

Role coverage disappears from scoring entirely (roles are now model-assigned labels, not
ranking features). This costs some recall ordering; the explorer model compensates because it
now sees candidates inline and can keep exploring when the top of the list looks wrong.

## Token Accounting (why this nets out)

Per `explore_code` call, V1 (candidate-followup) vs V2:

| Cost center                     | V1                          | V2                         |
| ------------------------------- | --------------------------- | -------------------------- |
| Sub-agent LLM invocations       | up to 6 sequential          | 1 (+<=2 continuations)     |
| Candidate packet resends        | up to 3 x 3-8k tokens       | 0 (inline IDs)             |
| Report into main context        | ~2k tokens, 4x redundant    | ~600 tokens, single-form   |
| Per-call policy text (uncached) | 400-600 chars every call    | 0                          |
| Padding files                   | always 5 primary            | only what the model picked |
| Rediscovery after distrust      | common (unverifiable facts) | reduced (verbatim quotes)  |

The dominant lever is the last row. A report the main model trusts replaces 5-20 broad main
reads; a report it distrusts is pure overhead on top of them. Quotes plus validated facts are
the trust mechanism; everything else is supporting cost reduction.

## Implementation Sequence

### Phase A: submit_report + inline IDs (the collapse)

- Add stable candidate-ID assignment at observation time; annotate rendered tool results.
- Add `submit_report` tool with the V2 selection schema; force completion with one nudge.
- Delete the selection pass, both follow-up passes, the minimum-evidence pass, and JSON
  scraping. Wire the validator + continuation loop.
- Keep the deterministic report builder as the model-failure fallback only.
- Rebaseline `explore_code_subagent.spec.ts` around invariants: unknown IDs cannot render,
  quotes must match evidence, budgets hold, downgrades-only validation, fallback works. Do not
  assert specific winners.

Acceptance: one streamText conversation per explore call on the happy path; selection arrives
as a validated tool call; all existing invariant tests pass.

### Phase B: intent argument + word-list deletion

- Add `intent` to `exploreCodeSchema`, tool description, main prompt guidance, and cache key.
- Delete `inferTaskIntent`, morphology/action/low-signal lists, role regexes, `target` role,
  product-shaped support paths.
- Reduce `buildCandidate` to structural features.
- Add the **domain-literal guard test**: production explorer files must not contain benchmark
  repo names, product nouns, or task vocabulary (`scene`, `channel`, `invoice`, `excalidraw`,
  `mattermost`, ...). This test would have caught today's residue.

Acceptance: guard test passes; no intent regexes remain; scoring is explainable from structure.

### Phase C: report V2 + prompt tightening

- New renderer (single representation, 2.5k budget, quotes, slim JSON block).
- Update `explore_code.ts` cache-stat extraction for the slim JSON block.
- Rewrite tool description and `local_agent_prompt.ts` guidance for the V2 contract; move all
  policy there.

Acceptance: report chars p50 <= 2.5k; every path renders once; zero imperative sentences in
report bodies.

### Phase D: measurement

- Benchmark arms: `explore-v1` (current candidate-followup) vs `explore-v2`, plus baseline.
- Headline metric: **main-model tool calls after a high/medium report** (broad grep/list_files
  count, read_file calls outside read targets), on a held-out repo split (>=8 tasks from repos
  never used to write any heuristic), repeats >= 3.
- Secondary: main uncached input p50, report tokens, sub-agent invocations and elapsed time,
  `fact_unverified` rate, continuation-round distribution, answer quality rubric.

Acceptance: held-out quality >= V1; main uncached input and post-report broad calls decrease;
no production diff reintroduces domain literals.

## Risks

- **Cheap model can't drive a forced tool call reliably.** Mitigation: one nudge retry, then
  deterministic fallback (already exists). Measure the fallback rate; if it exceeds ~10%,
  revisit model choice for the sub-agent before adding orchestration back.
- **Quote validation is too strict** (whitespace/truncation mismatches drop real facts).
  Normalize aggressively (collapse whitespace, strip line numbers), match against untruncated
  observation text, and track `fact_unverified` rate before tightening further.
- **Deleting word lists drops benchmark recall.** Expected and acceptable in-sample; the
  held-out split is the metric that matters. Do not add compensating vocabulary back.
- **Inline ID annotations confuse the explorer model.** Keep them terse (`[c7]`) and explain
  them once in the system prompt; verify with a few manual traces before benchmarking.
- **Spec rebaseline is large.** It is — but it is the last rebaseline of this size if the tests
  assert invariants instead of winners.

## Acceptance Criteria

- One sub-agent conversation per explore call; selection is a schema-validated tool call.
- The value model still cannot author paths or ranges; additionally, every rendered fact is
  backed by a verbatim observed quote.
- No morphology tables, action/intent word lists, closed role enums, or domain literals in
  production explorer code (enforced by a guard test).
- Reports are single-representation, <=2.5k chars, with policy text only in cached context.
- Held-out benchmark: post-report main broad-search calls and main uncached input decrease
  versus V1 without quality regression, across >=3 repeats.

Explore V2: Single-Conversation Explorer With a Schema-Bound Finish

Explore V2: Single-Conversation Explorer With a Schema-Bound Finish

Summary

Goals

Non-Goals

Current State (what to keep, what to delete)

Keep

Delete

Known bugs to fix in passing

Architecture

1. Candidate IDs injected inline

2. submit_report as a forced tool call

3. Intent comes from the caller

4. Adaptive continuation instead of follow-up passes

5. Validator (the only deterministic policy)

6. Report format: one representation

2. `submit_report` as a forced tool call