Operation Log — ACS

• 2026-07-12 (handoff) — No code changes; a walkthrough + doc-cleanup session. Explained the shipped classifier architecture to Arijit end-to-end (agent roster, synchronous call flow, decision logic with fire/no-fire examples), reading instructions_classifier.md directly rather than recalling from memory. Archived the stale 2026-07-09 tool-call-revert plan doc to docs/plans/archive/ with a superseded-note. Deleted and rewrote docs/plans/2026-07-10-reconciled-handoff-architecture-build.md in place — its Track A/B decision framing never actually shipped; new version documents the classifier design as-built, then carries forward only the AC2-port task list, correctly scoped to reuse the classifier architecture (3 decisions still pending Arijit: agent roster collapse, browser-direct vs backend-mediated, dead-code disposition). Ran the brainstorming skill on the previously-unscoped "sample-questions UI hint" — root-caused the real problem (chips exist on the empty state but read as visually inert, only a hover state reveals they're clickable) and wrote + got approval on a small spec: docs/superpowers/specs/2026-07-10-empty-state-sample-question-nudge-design.md (copy label + promoted resting-style chip styling, empty-state only — NOT built yet, next session's first task). Collected 3 decisions from Arijit on queued open items: mandate-guard.sh gets a real hook redesign (not a settings.json allow-rule) — not started; AC2 port scoped now, not deferred — 3 decisions still open before code; 100-Q eval + judge calibration explicitly deferred again. Commits: fc2b9b8, 24887d0.
• 2026-07-10 (handoff) — Session closed for real. Zero new work since the last persist. Classifier fix (860ed3e), empty-answer fix (be576e7), and the real numbered-list fix (c3ad2b9) all remain shipped, deployed, and live-verified. Next session: SESSION.md is the resume artifact.
• 2026-07-10 (session close) — No new work since the prior persist this session. Session closing clean: classifier fix (860ed3e), empty-answer fix (be576e7), and the real numbered-list fix (c3ad2b9) all remain shipped, deployed, and live-verified. Nothing pending except the open items already logged (sample-questions UI ask, stale plan doc, AC2 port, mandate-guard.sh).
• 2026-07-10 (post-midnight) — Shipped Track A: fixed the real offer-caching-race bug with a new dedicated ACS-classifier-neural agent (dbb4faa9-e917-4be9-b8ee-6dfd9a81daef), replacing Agent Studio's racy async config.suggestions job entirely. Full development-loop build (A1-A9, FULL path — Risk STRIDE + Architecture Review both run since this touches a live production agent ID). Architecture Review caught 2 Criticals before Build: a query-provenance/injection gap in the (still-being-considered) orchestrator design, and no way to give an agent zero search tools. Pure-orchestrator design (validated the night before via real E2E, 10/10) was then deliberately KILLED for ACS — once the classifier existed, the orchestrator's routing decision became provably redundant, adding 2 network hops for zero behavior change. Kept fully specced, shelved for AC2's port. Live acceptance-gate probe (7 pairs against real -dev agents) caught a real classifier over-triggering bug (false SPECIALIST offers on ambiguous design questions) BEFORE the live flip — fixed the prompt's "if unsure, default to SPECIALIST" tie-break, re-verified 7/7. Merged (squash), pushed, deployed, browser-verified with a real repeated question (offer survived identically both times — the actual bug this build exists to fix). Commit: 860ed3e (main fix).
• 2026-07-10 (post-midnight, continued) — 3 more real bugs, all caught by Arijit actually using the live deployed site, none caught by the automated test suite: (1) the classifier still fired and offered a deep-dive on an EMPTY Generic answer (the known ~1-in-8 platform flake) — fixed by skipping the classifier call when the answer is blank (be576e7). (2) A numbered-list rendering bug — the markdown-lite renderer only recognized bullet lines, so numbered lists fell through to a plain paragraph with no list styling. First fix attempt (ca19f83) added a 4px spacing change — Arijit correctly rejected it as invisible/not a real fix. (3) The REAL root cause, found by checking document.querySelector('ol') on the live page and getting null: the renderer classified a WHOLE text block as list-or-prose, so one intro sentence sharing a block with a numbered list (the normal shape — no blank line before "1. Do X") sent the entire block, list included, to the plain-paragraph path. Fixed by grouping lines into same-kind consecutive runs instead of classifying the whole block (c3ad2b9), verified via a real DOM query on production, not a screenshot. Also hit a self-inflicted persist tooling issue: bundling a git-commit shell command with a git-push shell command in the same invocation caused the push guard to block the ENTIRE command including the commit, twice — mitigation is one action per shell call, never combine a commit with a push/deploy in the same invocation.
• 2026-07-10 (very early) — Arijit halted further building on the revert-to-tool-call plan (written 20:34, 0% executed) and tonight's separate pure-orchestrator spike (faked test data, flagged as unproven). Ordered a real E2E test on real data before any more code. Interviewed him: pure-orchestrator confirmed authoritative, live -dev API OK, full T1-T6 suite run at once. Built + ran scripts/spikes/agent-tool-handoff/e2e-orchestrator-validation.mjs. Result: 10/10 real offer-signal precision, 10/10 real orchestrator-chaining-correctness (test script's own scoring had a null===false bug that under-reported the second number as 5/10 — architecture is fine, the counter was wrong). T6 (browser) not run. No production code touched. Files: index.md, wiki/dev-log.md, wiki/open-questions.md (+3), tasks.md (new), memory (2 new + session_pointer), SESSION.md.
• 2026-07-09 (late evening) — Shipped + deployed 2 UI fixes (6e430f5): real bullet-list markdown rendering, source-group count badge always clickable (fold/unfold whole group). Verified live in served bundle. Wrote exhaustive porting playbook distilling both build days (3 handoff-mechanism generations, 17 gotchas, test coverage, env vars, deploy checklist) for reuse on Algolia-Central2 — wiki/porting-playbook-to-algolia-contentsearch.md, mirrored to Algolia-Central2/docs/porting-playbook-from-acs.md. Arijit requested a spoken-word ("arijit-approved") unlock for mandate-guard.sh in place of the file-token dance; design written, not applied (self-mod of the assistant's own permission hook correctly refused by the auto-mode classifier — Arijit needs to paste it in himself).
• 2026-07-09 (persist, evening) — Full build shipped + live incident closed. Suggestions migration (B0-B9) deployed to main+prod. Live incident: gemini-2.5-flash-lite deprecated by provider, switched to gemini-2.5-flash. 3 more fixes found+shipped during incident response: Generic brevity restored, recap-previous-turn bug fixed, poisoned sample-question reworded. mandate-guard.sh/classifier conflict surfaced, unresolved. Commits: 111c8fd, 225b04b, 0958e32, 64c48fc, d7a999c. Files: index.md (compiled-truth zone rewritten), wiki/log.md (vault), MEMORY.md + 2 new memory files, SESSION.md.
• 2026-07-09 — Recorded ADR: dropped the 07-08 tool-call handoff design before building it, replaced with Agent Studio's native config.suggestions feature as the offer-detection signal (app already orchestrates both agents directly, no model-driven cross-agent call needed). Collapses the old separate suggestions/follow-up job into the same mechanism. Confirmed multi-turn continuity needs no new platform capability. Files: wiki/decisions/2026-07-09-drop-tool-call-handoff-use-native-suggestions.md, wiki/dev-log.md (new), wiki/open-questions.md (+3 questions), index.md.
• 2026-07-08 (session close #2) — Live bug fixed (Technical "no response" flake, auto-retry added); installed real algoliasearch+react-instantsearch deps as react-instantsearch migration step 1; researched <Chat> API for real (agentId is single, no multi-agent — tools prop maps to our tool-call mechanism); redesigned follow-up-question instructions on both agents (specific/varied, not templated) + gave Technical a follow-up mechanism it never had + fixed a frontend bug that would have leaked raw tokens. Verified live. Pushed fac01d5.
• 2026-07-08 (session close #1) — Replaced **HANDOFF:technical** text sentinel with a real live-tested consult_technical_specialist client_side tool call (Agent Studio). Redeployed both agents. Built orchestratorRunner.ts eval harness + 120-Q bank on @lab/judge; found/fixed 3 real bugs (frontmatter-in-grounding-text, Generic under-searching on comparisons, res.text() cache-poisoning). 20-Q pilot mean 8.02→8.61. Full 100-Q run not completed. Committed + pushed to origin/spike/agent-to-agent-tool. Arijit flagged: entire chat client is hand-built, zero react-instantsearch/algoliasearch deps — top item for next session. Full detail: index.md timeline.
• 2026-07-01 — Project initialized. Fork of Algolia-Central2 for Adobe (prospect). Vault folder created; repo CLAUDE.md written at ~/Dropbox/AI-Development/RAG/ALgolia-Central-Spectrum. Namespace ACS- locked. Corpus source algolia-central.vercel.app/docs (to confirm). DocSearch MCP flagged as candidate ingest/retrieval layer. Open questions logged.
• 2026-07-01 (session 1) — Corpus complete: 502 recs neural in ACS_SPECTRUM_MULTI (both Adobe sources fully crawled + react-aria). Built 2 agents (ACS-generic-neural + ACS-technical-neural, verified). Shipped algolia-content-fetch skill (3 engines). Ported judge harness (runs e2e, scoring bug open). NEXT = UI. All committed on ACS repo. /persist + /handoff at close.
• 2026-07-02/03 — Chat UI built end-to-end (Vite/React/TS): human-gated deep-dive, real agent-generated discovery follow-up, grouped sample questions, light Algolia design system (Sora, Nebula Blue, cobrand Adobe×Algolia). Corpus V3 topped up 95→144 (358 total). Committed + pushed to GitHub (https://github.com/arijitchowdhury80/algolia-central-spectrum).
• 2026-07-03 (night) — Repo docs shipped (README with inline Mermaid diagrams, docs/ARCHITECTURE.md code map, Excalidraw system/turn-flow diagrams) + GitHub description/topics set — repo had zero metadata before this. Vercel deploy fixed and verified live: root cause was no vercel.json (app lives in web/, not repo root); added it, set VITE_ALGOLIA_APP_ID/VITE_ALGOLIA_SEARCH_API_KEY via vercel env add (extracted search key from the built bundle, validated against the live index), redeployed via vercel --prod. https://algolia-central-spectrum.vercel.app confirmed HTTP 200 with correct title and inlined config. Pushed as 973add9. Unverified: Agent Studio CORS from the vercel.app origin (needs a real browser query to confirm).
• 2026-07-03 (late-night, fabbd07, v0.3.0) — Surfaced the grounding judge in the chat: Confidence chip on each answer → JudgeDrawer (3-judge accordions Skeptic/Referee/Advocate + dynamic dims + rationale); retired the passive right-rail panel. Wired to the HOSTED VPS judge (ac2-lab-backend → judge.contentengagement.info, x-lab-key). Fixed the grounding hard-gate falsely capping live scores to 3.0 at the backend source (liveJudge.ts toVerdict: cap only on a real, shown contradiction — not the 1-round recurrence artifact / hidden unverifiable flags); rebuilt the VPS container. Browser-verified live: the ComboBox answer that read 3.0/UNSUPPORTED now reads 8.9/GROUNDED; a fabricated answer still caps to 0 with the fake claims listed. Made SourcePills counts/"+N" expandable to verify each source. Patched the base UI template + judge docs (playbook §7, judge/README-artifact.md, BUILD-STATUS) for turnkey reuse on future instances.