@event4u/agent-config 14.12.0 → 14.14.0

package/src/scripts/council_cli.ts

@@ -26,6 +26,7 @@ import {
     bundle_roadmap,
 } from './ai_council/bundler.js';
 import { select_chairman } from './ai_council/chairman.js';
+import { warnIfRecounciled } from './ai_council/recouncil_guard.js';
 import { jaccardSimilarity } from './_lib/text_similarity.js';
 import type { ChairmanCandidate } from './ai_council/chairman.js';
 import { synthesis_template } from './ai_council/prompts.js';
@@ -137,19 +138,23 @@ import {
     run_debate,
     run_peer_review,
 } from './ai_council/orchestrator.js';
+import { serializeResponses } from './ai_council/response_record.js';
 import {
     type PriceTable,
     downgrade_coupling,
-    estimate_cost,
+    allSeatsNonBillable,
     estimate_input_tokens,
     load_prices,
     prices_file_for,
+    sumBillableCost,
 } from './ai_council/pricing.js';
 import { detect_project_context } from './ai_council/project_context.js';
 import {
     DecisionReplayInputs,
     render_decision_replay,
 } from './ai_council/replay.js';
+import type { FindingsExtraction } from './ai_council/consensus.js';
+import { harvest_inline_findings, inlineFindingsActive } from './ai_council/inline_findings.js';
 import {
     Finding as ConsensusFinding,
     FindingScore as ConsensusFindingScore,
@@ -471,8 +476,6 @@ function load_settings(
     return settings;
 }
 
-
-
 // ── member construction ─────────────────────────────────────────────
 
 interface BuildMembersOptions {
@@ -1186,6 +1189,7 @@ function _maybe_run_consensus(
     table: PriceTable,
     project: unknown,
     args: Args,
+    inline_extractions: ReadonlyMap<string, FindingsExtraction> | null = null,
 ): ConsensusResult | null {
     const cs = (ai_cfg['consensus_scoring'] as Dict) || {};
     if (!_pyBool(cs['enabled'])) {
@@ -1196,6 +1200,7 @@ function _maybe_run_consensus(
         return null;
     }
     return run_consensus_scoring(members, responses, {
+        inline_extractions,
         budget,
         table,
         project: project as never,
@@ -1230,8 +1235,8 @@ function _serialise_consensus(consensus: ConsensusResult): Dict {
             reason: s.reason,
         })),
         metadata,
-        extraction_responses: _serialise_responses(consensus.extraction_responses),
-        scoring_responses: _serialise_responses(consensus.scoring_responses),
+        extraction_responses: serializeResponses(consensus.extraction_responses),
+        scoring_responses: serializeResponses(consensus.scoring_responses),
     };
 }
 
@@ -1478,7 +1483,7 @@ function _maybe_run_peer_review(
 
 function _serialise_peer_review(peer_review: PeerReviewResult): Dict {
     return {
-        responses: _serialise_responses(peer_review.responses),
+        responses: serializeResponses(peer_review.responses),
         label_to_source: _mapToObject(peer_review.label_to_source),
         persona_labels: _mapToObject(peer_review.persona_labels),
     };
@@ -1773,28 +1778,6 @@ function _emit_debate_estimate(
     return 0;
 }
 
-function _serialise_responses(responses: CouncilResponse[]): Dict[] {
-    const out: Dict[] = [];
-    for (const r of responses) {
-        const metadata: Dict = {};
-        for (const [k, v] of Object.entries(r.metadata || {})) {
-            metadata[k] = String(v);
-        }
-        out.push({
-            provider: r.provider,
-            model: r.model,
-            text: r.text,
-            input_tokens: r.input_tokens,
-            output_tokens: r.output_tokens,
-            cache_creation_input_tokens: r.cache_creation_input_tokens,
-            cache_read_input_tokens: r.cache_read_input_tokens,
-            latency_ms: r.latency_ms,
-            error: r.error,
-            metadata,
-        });
-    }
-    return out;
-}
 
 function _deserialise_responses(items: Dict[]): CouncilResponse[] {
     const out: CouncilResponse[] = [];
@@ -2503,6 +2486,7 @@ function cmd_run(
     const [extra_calls, extra_usd] = _consensus_cost_delta(ai_cfg, question.mode, estimates, billable.length);
     const [ch_calls, ch_usd] = _chairman_cost_delta(ai_cfg, estimates);
     const [pr_extra_calls, pr_extra_usd] = _peer_review_cost_delta(ai_cfg, args, estimates, billable.length);
+    warnIfRecounciled(REPO_ROOT, artefact, question.user_prompt, members, _resolve_rounds(args, ai_cfg), _stdout);
     if (!args.confirm) _stdout('council:run · DRY PASS — estimate only, no seat is contacted. Add --confirm to run.\n');
     _stdout(
         `council:run · mode=${question.mode} · members=${members.length} ` +
@@ -2623,6 +2607,9 @@ function cmd_run(
         rounds,
         advisor_plans,
         stance_tally: stance_tally_on,
+        // Phase 1B — the same predicate the harvest below reads, so a member is asked
+        // for the block exactly when the consensus round intends to read one.
+        inline_findings: inlineFindingsActive(ai_cfg, question.mode),
         member_prompt_suffix,
         no_project_context_members,
         // Mid-flight cli→api fallback (ai-council-config.md § failure-class-
@@ -2706,6 +2693,9 @@ function cmd_run(
     // this is always additive, never a fabricated decision.
     const handoff: HandoffEnvelope = buildHandoffFromStanceTally(stance_tally_result);
     const persona_labels = build_persona_labels(advisor_plans, billable);
+    // Phase 1B — BEFORE peer review, chairman synthesis and rendering read
+    // `responses`; the block is scaffolding, not argument. Off → null, nothing touched.
+    const inline_extractions = inlineFindingsActive(ai_cfg, question.mode) ? harvest_inline_findings(members, responses) : null;
     const peer_review = _maybe_run_peer_review(
         ai_cfg,
         args,
@@ -2717,7 +2707,7 @@ function cmd_run(
         project,
         { persona_labels },
     );
-    const consensus = _maybe_run_consensus(ai_cfg, question, members, responses, budget, table, project, args);
+    const consensus = _maybe_run_consensus(ai_cfg, question, members, responses, budget, table, project, args, inline_extractions);
     // Rendered attendance, re-derived after the parser. The event does not move.
     quorum_out.result = annotateRenderedQuorum(quorum_out.result, consensus?.parse_outcomes);
     const chairman = _maybe_run_chairman(
@@ -2761,13 +2751,10 @@ function cmd_run(
         all_responses.push(chairman.response);
     }
     all_responses.push(...stance_repairs);
-    for (const r of all_responses) {
-        if (r.error) {
-            continue;
-        }
-        const ce = estimate_cost(r.provider, r.model, r.input_tokens, r.output_tokens, table);
-        actual_total += ce.input_usd + ce.output_usd;
-    }
+    // Billable-aware: a subscription-CLI seat spent nothing, so pricing it at
+    // API rates and calling the figure "actual" is a false statement about
+    // money. See pricing.ts § billable-aware aggregation.
+    actual_total += sumBillableCost(all_responses, table);
     const payload: Dict = {
         schema_version: SCHEMA_VERSION,
         mode: question.mode,
@@ -2784,7 +2771,7 @@ function cmd_run(
         stances: stances_on,
         blind_chairman: blind_chairman_on,
         chairman_fields: chairman_fields_on,
-        responses: _serialise_responses(responses),
+        responses: serializeResponses(responses),
     };
     if (peer_review !== null) {
         payload['peer_review'] = _serialise_peer_review(peer_review);
@@ -2819,9 +2806,16 @@ function cmd_run(
     });
     fs.mkdirSync(path.dirname(_resolveTarget(out_path)), { recursive: true });
     fs.writeFileSync(_resolveTarget(out_path), _jsonDumpsIndent2(payload) + '\n', { encoding: 'utf-8' });
+    // The trailing note is not decoration. Before 2026-08-27 this line read
+    // `actual $0.1055` on a run whose every seat was subscription-authed, two
+    // lines below a `TOTAL: $0.0000` from the pre-run path — and an operator
+    // reading the tail of the output saw only the wrong one. When every seat is
+    // non-billable the figure is now 0 and says why, so the two lines agree.
+    const _all_subscription = allSeatsNonBillable(all_responses);
     _stdout(
         `\ncouncil:run · wrote ${out_path} ` +
-            `(estimated $${_pyFixed(estimated_total, 4)} / actual $${_pyFixed(actual_total, 4)})\n`,
+            `(estimated $${_pyFixed(estimated_total, 4)} / spent $${_pyFixed(actual_total, 4)}` +
+            `${_all_subscription ? ' — all seats subscription-authed, nothing billed' : ''})\n`,
     );
     const replay_path = _maybe_write_decision_replay({
         ai_cfg,
@@ -2863,14 +2857,9 @@ function _write_debate_round(
     },
 ): string {
     fs.mkdirSync(_resolveTarget(out_dir), { recursive: true });
-    let actual_total = 0.0;
-    for (const r of responses) {
-        if (r.error) {
-            continue;
-        }
-        const ce = estimate_cost(r.provider, r.model, r.input_tokens, r.output_tokens, opts.table);
-        actual_total += ce.input_usd + ce.output_usd;
-    }
+    // Billable-aware, same reason as the sibling site above: a subscription
+    // seat contributes nothing to `cost_usd_actual`.
+    const actual_total = sumBillableCost(responses, opts.table);
     const payload: Dict = {
         schema_version: SCHEMA_VERSION,
         mode: opts.question.mode,
@@ -2884,7 +2873,7 @@ function _write_debate_round(
         rounds: 1,
         cost_usd_actual: _pyRound(actual_total, 6),
         prompt_cache_round_gap_ms: opts.cache_gap_ms_since_previous_round ?? null,
-        responses: _serialise_responses(responses),
+        responses: serializeResponses(responses),
     };
     const out_path = path.join(_resolveTarget(out_dir), _debate_round_filename(round_number));
     fs.writeFileSync(out_path, _jsonDumpsIndent2(payload) + '\n', { encoding: 'utf-8' });
@@ -3220,13 +3209,7 @@ function cmd_debate(
 
     let actual_total = 0.0;
     for (const rnd of [...all_rounds, restate_responses]) {
-        for (const r of rnd) {
-            if (r.error) {
-                continue;
-            }
-            const ce = estimate_cost(r.provider, r.model, r.input_tokens, r.output_tokens, table);
-            actual_total += ce.input_usd + ce.output_usd;
-        }
+        actual_total += sumBillableCost(rnd, table);
     }
     _stdout(
         `\ncouncil:debate · ${all_rounds.length} round(s) complete · ` +

package/src/scripts/dispatch_r2_reviewer.ts

@@ -54,6 +54,7 @@ import { fileURLToPath, pathToFileURL } from 'node:url';
 // time, and each CLI entry guard only fires for its own argv[1].
 import { isAcceptanceCriteriaHeading } from './_lib/ac_heading.js';
 import { gitEnv } from './_lib/git_env.js';
+import { loadDenyPatterns, redactSourceShape, redactSourceTokens, writeRedacted } from './_lib/source_redact.js';
 import { completionReviewDisabled } from './_lib/planning_settings.js';
 import { artifactRelevance } from './check_completion_review.js';
 
@@ -1251,17 +1252,41 @@ function runDispatch(args: Args): number {
         acHash: hashes.ac_hash,
         dispatched,
     });
+    // Phase 3.4 of road-to-source-silence: redact deny-set hits AT WRITE TIME.
+    // This package's confidentiality gate carried a blanket `skip_paths`
+    // exemption over `*.review-input/diff.patch`, which did not stop the leak —
+    // it stopped the gate from seeing it. The generator is the only writer of
+    // these files, so it is the only place the content can be fixed instead of
+    // exempted. Redaction is line-local: hunks, paths and context are
+    // untouched, so the evidence chain a reviewer reads stays intact.
+    //
+    // ORDERING IS LOAD-BEARING. `prompt_hash` must re-derive from the bytes on
+    // disk (`check_review_prompt_binding` re-hashes `prompt.md`), so the prompt
+    // is redacted BEFORE it is hashed. Hashing the pre-redaction text and
+    // writing the post-redaction text would make every new artefact fail its
+    // own binding gate.
+    const denyPatterns = loadDenyPatterns();
+    let redactions = 0;
+    const promptTokens = redactSourceTokens(promptText, denyPatterns);
+    // Shape redaction runs on the prompt too, and BEFORE the hash below, for the
+    // same reason the deny pass does: `check_review_prompt_binding` re-derives
+    // prompt_hash from prompt.md on disk, so hashing anything other than the
+    // exact bytes written would make every new artefact fail its own binding.
+    const promptRedacted = redactSourceShape(promptTokens.text);
+    redactions += promptTokens.count + promptRedacted.count;
+
     const skeleton = findingsSkeleton({
         slug,
         headSha,
         scopeHash: hashes.scope_hash,
         reviewedDate,
         manifest,
-        promptHash: sha256(promptText),
+        promptHash: sha256(promptRedacted.text),
     });
 
     fs.mkdirSync(inputDirAbs, { recursive: true });
-    fs.writeFileSync(path.join(inputDirAbs, 'diff.patch'), scopeDiffText, 'utf-8');
+
+    redactions += writeRedacted(path.join(inputDirAbs, 'diff.patch'), scopeDiffText, denyPatterns);
     if (roadmapText !== null) {
         // The snapshot lands under agents/evidence/, which check_references
         // walks, while the live roadmap layer is deliberately excluded from
@@ -1273,10 +1298,26 @@ function runDispatch(args: Args): number {
         const snapshotHeader =
             '<!-- check-refs: skip -->\n' +
             '<!-- verbatim roadmap snapshot for the R2 reviewer; the live roadmap layer is excluded from check_references, and a snapshot must not fail a gate its source is exempt from -->\n';
-        fs.writeFileSync(path.join(inputDirAbs, 'roadmap.md'), snapshotHeader + roadmapText, 'utf-8');
+        redactions += writeRedacted(
+            path.join(inputDirAbs, 'roadmap.md'),
+            snapshotHeader + roadmapText,
+            denyPatterns,
+        );
+    }
+    redactions += writeRedacted(
+        path.join(inputDirAbs, 'acceptance-criteria.md'),
+        acText ?? '',
+        denyPatterns,
+    );
+    fs.writeFileSync(path.join(inputDirAbs, 'prompt.md'), promptRedacted.text, 'utf-8');
+    if (redactions > 0) {
+        // Never redact silently: a changed artefact nobody was told about is
+        // the failure mode this whole programme is about.
+        process.stderr.write(
+            `dispatch_r2_reviewer: redacted ${redactions} source-confidentiality hit(s) ` +
+                `from the review-input snapshot (marker: [REDACTED:src-conf]).\n`,
+        );
     }
-    fs.writeFileSync(path.join(inputDirAbs, 'acceptance-criteria.md'), acText ?? '', 'utf-8');
-    fs.writeFileSync(path.join(inputDirAbs, 'prompt.md'), promptText, 'utf-8');
     fs.writeFileSync(findingsAbs, skeleton, 'utf-8');
 
     const inputDirRel = path.join(args.outDir, `${slug}.review-input`);