feat(agents): proactive "render UI by default" directive for MAF + Copilot agents (genUI Phase 2) - #190
Merged
Conversation
BreathingCharacter default sprite scale 1.35x -> 1.8x of its box (the frames carry heavy transparent padding, so the character reads much bigger without cropping), and boxes now nearly fill their fixed columns: picker cards 52->60px, agents grid tiles 56->62px, agent settings header 76->92px. Card and column dimensions unchanged. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01K9rzBFbJiEtwSyq8u7sWWt
…mework-review-qvimev
…th durable script store Give every MAF agent the Copilot SDK as a CAPABILITY (not a peer agent), per the framework review's dual-runtime verdict: - code_task(task): one-shot, 600s-bounded Copilot coding session in the calling agent's own workspace (orchestrator/code_session.py). Manifest- first contract: read agent-data/SCRIPTS.md, edit scripts in place under agent-data/scripts/, update the manifest. BYOK via the gateway /v1. - run_script(path, args?): zero-LLM re-execution of a saved .py/.sh script — cheap reuse of capability built in previous sessions. Durability: scripts live under blob-store-backed agent-data/ (Postgres authoritative), surviving restarts, redeploys, and volume wipes. Both tools finish with an explicit sweep that mirrors native-written files (Copilot CLI + script self-writes bypass the write_artifact mirror) into the store — code_task sweeps even when the session failed. Execution hygiene: workspace-jailed path resolution (fail closed on escape), secret-scrubbed subprocess env (allowlist + deny pattern — no tokens/keys/DB URLs reach script code), wall-clock timeouts, capped output. Registered open-world in TOOL_ANNOTATIONS; rides the permission gate like every injected tool. Wiring: injected into all agent shapes, added to the guaranteed core tool floor, documented in the full + compact addendum. 24 new unit tests; spec at ai-company-brain/specs/agent_coding_skill.md. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01K9rzBFbJiEtwSyq8u7sWWt
…ed skill fixes flow through the approval pipeline The agent's chat workspace IS its persistent git clone, so the coding skill and the existing self-mutation machinery are one system: - Coding-session contract now teaches TWO script homes: agent-data/ scripts/ (workspace scripts, blob-durable, no approval) vs git-tracked repo source (skills/*/scripts/, agents.py — the agent's built-in skills). Repo skills are edited IN PLACE and committed locally, never pushed — the executor's post-run commit scan registers every local commit as a pending_commit inbox row and the pre-push hook blocks direct pushes, same as failure-driven mutation and loader auto-sync. - Fail-safe in code_task (_commit_repo_changes): if a session edits tracked source and forgets to commit, code_task commits the residue — the loader's next _pull_latest stash-drops/hard-resets uncommitted tracked changes, so an uncommitted skill fix would be silently destroyed. git add -A respects the loader-managed .gitignore, so agent-data/, inputs/, outputs/ never leak into a commit. - Addendum + tool doc tell agents that a misbehaving built-in skill is fixed via code_task and goes live after human approval. - Spec: §7 harmonisation — script taxonomy + the three mutation paths (failure-driven, chat-driven, promotion via loader auto-sync) all converging on pending_commit → approve → push. 5 new tests (mini-clone fixture): fail-safe commit + clean tree, no noise commits, non-git workspaces, contract/addendum coverage. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01K9rzBFbJiEtwSyq8u7sWWt
The sweep compared file mtimes against a time.time() cutoff, but kernel file timestamps come from a coarse clock that can lag time.time() by a few ms — a file written immediately after the cutoff capture could carry an mtime just BEFORE it and be silently skipped (a very fast code_task write would then miss the blob-store mirror; surfaced as a CI-only flake in test_code_task_failure_still_sweeps). Slacken the cutoff by 2s; re-mirroring a file modified moments before the run is harmless because the mirror is an idempotent write-through. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01K9rzBFbJiEtwSyq8u7sWWt
…st_integrations discoverability Scripts often exist precisely to call an external service the platform holds credentials for — but the scrubbed subprocess env blocked even the integrations an agent legitimately declared, while blanket pass-through would leak the whole secret surface. Apply the platform's existing rule to the coding skill: an agent's scripts get EXACTLY the integrations that agent declared in config.json — nothing else. - acb_skills.integrations.FIELD_TO_ENV: canonical service→env-var mapping, now the single source of truth shared by the executor's run-scoped env injection and the script env (also fixes google-maps, which was missing from the executor's private copy and therefore never env-injected). env_var_names() helper. - Executor publishes each run's resolved integration names (+ warnings) into the tool context; sub-agents get their own declared set, not the parent's (saved/restored around delegation). - code_tools._script_env grants the base allowlist PLUS exactly the declared services' env vars; gateway master key, DB URL, and undeclared integrations stay scrubbed (proven by live subprocess test). - New injected core-floor tool list_integrations(): resolved services + env-var NAMES scripts may read + declared-but-unavailable reasons — never credential values. - code_task session prompt names the available vars (names only); the harness forbids hard-coding, printing, or persisting credential values. Addendum documents the workflow. - Spec §8; 9 new tests (34 total in test_code_tools.py). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01K9rzBFbJiEtwSyq8u7sWWt
…mework-review-qvimev
…times (genUI Phase 2) Agents could create generative UI but weren't reaching for it. Two root causes, both on the prompting side (emit path + frontend render are already runtime-agnostic): 1. The "reach for it eagerly" guidance lived only in the tool docstring (weak signal) and, at system-prompt level, buried mid-addendum. No answer-time nudge to convert a data/status/comparison/choice reply into UI. 2. Native MAF agents got NONE of it: _build_injected_tools_addendum (which carries the genUI guidance + design.md) is applied only on the Copilot _tools branch. email-assistant explicitly scopes in emit_generative_ui yet had zero system-prompt guidance to use it. Fix — _ui_first_directive: a static, byte-stable, tool-gated directive delivered to BOTH runtimes: - Template-first ordering (11 named templates → tree → custom HTML last) is the token-thrift + on-brand-by-construction lever; custom HTML carries the --cc-* token hook for the rare escape-hatch case. - Leads the Copilot addendum (full + compact) instead of being buried. - Parity fix: appended to native-MAF instructions (default_options and agent.tools shapes), marker-guarded for idempotency, compact variant for sub-agents. - Pick/set steered to formCard/optionPicker + "hitl":true. Also refreshes the addendum-drift eval to cover the coding-skill tools (run_script/code_task/list_integrations) that were injected-but-untracked. Tests: test_genui_proactive_directive.py (MAF instructions injection, idempotency, compact/full, byte-stability) + drift-eval assertions (directive leads both addenda, template-before-html, tool-gated). Spec generative_ui_2.md §7. Full suite green (1678 unit, 120 trajectory). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01K9rzBFbJiEtwSyq8u7sWWt
vjvarada
pushed a commit
that referenced
this pull request
Jul 23, 2026
…t (Granola pattern)
The single highest-leverage note-taker UX idea (research §3, spec §4
Tier-1 item 3): the user's attention during the meeting steers a far
better summary than the transcript alone.
- migration 97_note_taker_scratch.sql: meeting.scratch_notes TEXT (the
user's rough input notes, distinct from the generated meeting_note
output). Validated on a clean 00->97 chain + idempotency.
- summaries.py: generation reads scratch_notes and threads it into the
prompt via _scratch_block as EMPHASIS SIGNALS — expand the flagged
topics from the transcript, fix shorthand, ensure each is addressed,
but ground every fact in the transcript and never invent from the
notes. Works for both single-pass and map-reduce.
- meetings.py: PUT /meetings/{id}/scratch; scratch_notes surfaces on the
meeting detail.
- UI: a 'Jot notes as you go' textarea on the recording screen
(debounced autosave, flushed on stop) and a 'Your notes' editor on the
meeting detail (save on blur) — both feed the same field.
Rebased onto current main (picked up #189/#190). 3 new tests for the
scratch block (empty no-op, labelling + grounding guardrail + bound);
full suite 1694 passed; 19 /notes routes; frontend tsc clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016xdbRumGveUBFRsPh3Duu8
vjvarada
pushed a commit
that referenced
this pull request
Jul 23, 2026
…t (Granola pattern)
The single highest-leverage note-taker UX idea (research §3, spec §4
Tier-1 item 3): the user's attention during the meeting steers a far
better summary than the transcript alone.
- migration 97_note_taker_scratch.sql: meeting.scratch_notes TEXT (the
user's rough input notes, distinct from the generated meeting_note
output). Validated on a clean 00->97 chain + idempotency.
- summaries.py: generation reads scratch_notes and threads it into the
prompt via _scratch_block as EMPHASIS SIGNALS — expand the flagged
topics from the transcript, fix shorthand, ensure each is addressed,
but ground every fact in the transcript and never invent from the
notes. Works for both single-pass and map-reduce.
- meetings.py: PUT /meetings/{id}/scratch; scratch_notes surfaces on the
meeting detail.
- UI: a 'Jot notes as you go' textarea on the recording screen
(debounced autosave, flushed on stop) and a 'Your notes' editor on the
meeting detail (save on blur) — both feed the same field.
Rebased onto current main (picked up #189/#190). 3 new tests for the
scratch block (empty no-op, labelling + grounding guardrail + bound);
full suite 1694 passed; 19 /notes routes; frontend tsc clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016xdbRumGveUBFRsPh3Duu8
vjvarada
pushed a commit
that referenced
this pull request
Jul 23, 2026
…t (Granola pattern)
The single highest-leverage note-taker UX idea (research §3, spec §4
Tier-1 item 3): the user's attention during the meeting steers a far
better summary than the transcript alone.
- migration 97_note_taker_scratch.sql: meeting.scratch_notes TEXT (the
user's rough input notes, distinct from the generated meeting_note
output). Validated on a clean 00->97 chain + idempotency.
- summaries.py: generation reads scratch_notes and threads it into the
prompt via _scratch_block as EMPHASIS SIGNALS — expand the flagged
topics from the transcript, fix shorthand, ensure each is addressed,
but ground every fact in the transcript and never invent from the
notes. Works for both single-pass and map-reduce.
- meetings.py: PUT /meetings/{id}/scratch; scratch_notes surfaces on the
meeting detail.
- UI: a 'Jot notes as you go' textarea on the recording screen
(debounced autosave, flushed on stop) and a 'Your notes' editor on the
meeting detail (save on blur) — both feed the same field.
Rebased onto current main (picked up #189/#190). 3 new tests for the
scratch block (empty no-op, labelling + grounding guardrail + bound);
full suite 1694 passed; 19 /notes routes; frontend tsc clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016xdbRumGveUBFRsPh3Duu8
vjvarada
added a commit
that referenced
this pull request
Jul 23, 2026
… + Deepgram named speakers (#184) * feat(notes): slice 1 — notes generation, SSE progress, transcript↔notes UI Turns transcripts into grounded, structured meeting notes (spec §3.5): - templates.py: meeting-notes template as data (sections + instructions) → prompt compiler with the Meetily-derived grounding + anti-injection rules; markdown renderer for the canonical notes document. - summaries.py: notes generation on acb_llm (acompletion_with_fallback, strict JSON, transcript pinned as DATA). Single-pass for short transcripts, map-reduce for long ones (per-chunk cap is logged, never silent). Writes meeting.summary_md/json + meeting_note, and extracts draft action_item rows with segment_ids grounding (HITL approve→task is slice 2). Endpoints: POST /summarize, GET/PUT /note, GET /actions, GET /templates. Env-tunable model roles (NOTES_MEETING_SUMMARY_MODEL). - pipeline.py: auto-enqueues notes generation after transcription, so one upload yields transcript → notes with no second action. - events.py: per-meeting SSE progress stream (honest per-stage status, closes on terminal) + a dedicated Next stream proxy (the catch-all JSON proxy can't stream). - meeting detail: two-pane transcript↔notes (markdown via MarkdownMessage), live SSE-driven progress, and a confidence-ranked action-items rail. 9 new unit tests (template compiler, chunking, ref collection, markdown render); full suite 1647 passed. Gateway registers all /notes routes; frontend tsc clean. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016xdbRumGveUBFRsPh3Duu8 * feat(notes): in-browser recorder — live mic capture → chunked upload (slice 1) Completes slice 1's capture loop (R1: record a live conversation in-app): - recordings.py: chunked live-recording endpoints — POST /recordings/start (creates the row + empty file, marks the meeting 'recording'), POST /recordings/{id}/chunk?seq= (idempotent, gap-checked append; a MediaRecorder session's ordered timeslice blobs concatenate to a valid container), POST /recordings/{id}/complete (finalize + enqueue transcription, which auto-chains into notes generation). - recorder.ts: MeetingRecorder — getUserMedia (echo-cancel on), feature- detected mime (webm/opus, mp4 for Safari), MediaRecorder timeslice, and a sequential retry-with-backoff uploader that keeps unacked blobs queued in memory (the offline buffer for wifi blips) and flushes before complete. AnalyserNode drives a live level meter. - session/[id]: the 'studio, not form' recording screen (spec §5.1) — consent line, canvas waveform, large mono timer, pause/stop, backlog indicator, honest mic-permission errors. - Library 'Record' button (was disabled) now creates an in_person meeting and opens the session; 'Upload' demoted to secondary. - Notes proxy passes application/octet-stream bodies through byte-exact (the chunk uploads), alongside the existing multipart passthrough. Gateway registers all 12 /notes routes; full suite 1647 passed; frontend tsc clean on the new files. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016xdbRumGveUBFRsPh3Duu8 * feat(notes): action-item → task HITL loop-closure (slice 2) Turns draft action items into real GTD tasks — the Meeting→Task counterpart of the Email→Task capture (spec §3.9): - actions.py: POST /notes/actions/{id}/approve inserts a LOCAL gtd_items task (the store the task manager owns) with an origin={kind:meeting, meeting_id, action_item_id, segment_ids} provenance link, and records action_item.resulting_task_id for a two-way link. Idempotent (re-approve returns the existing task). /reject marks it rejected; /meetings/{id}/actions/approve-all approves every draft ≥ a confidence threshold. Every approval writes an audit_event. - Meeting detail action rail is now interactive triage (spec §5.5): Approve → Task / Reject per item, a confidence-ranked list, an 'Approve all ≥80%' bulk button, a success toast, and created items deep-link into /tasks. Read-only states for created/rejected. gtd_items columns (id/user_id/title/description from mig 48, origin from mig 65) verified. Gateway registers all 15 /notes routes; full suite 1647 passed; frontend tsc clean. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016xdbRumGveUBFRsPh3Duu8 * feat(notes): attendees + follow-up email loop-closure (slice 2) Closes the Notes→Email arc (spec §3.9): - migration 96_note_taker_attendees.sql: meeting.attendees JSONB [{name,email}] for external attendees (distinct from attendee_ids org person refs) — the recap recipient set. Validated on a clean PG16 chain run + idempotency; ORM mirror added. - meetings.py: attendees surface on the meeting detail; PUT /meetings/{id}/attendees replaces the list. - share.py: POST /meetings/{id}/share/email/draft LLM-drafts a recap (subject + body) from the notes (pinned as DATA), recipients from attendees; degrades to a raw-notes recap if the LLM is unavailable. Draft-only — nothing sends server-side. - UI: inline attendee chip editor on the meeting detail; a 'Follow-up' button opens a HITL compose modal (FollowupEmailModal) that fetches the draft + the user's email accounts, lets them edit to/subject/body and pick a sending account, and sends via the existing /email/send. Migration chain OK through 96; all 17 /notes routes register; full suite 1647 passed; frontend tsc clean. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016xdbRumGveUBFRsPh3Duu8 * feat(notes): ask-the-meeting — grounded Q&A over a transcript The first AI-intelligence feature (spec §4 Tier-1 item 2): ask a question about a meeting and get a transcript-grounded answer with clickable citations. - qa.py: POST /notes/meetings/{id}/ask reuses the summarization grounding discipline (transcript pinned as DATA, cite segment numbers, ignore embedded instructions). No precomputed embeddings needed — a meeting transcript is bounded, so it feeds relevant segments directly; when a transcript exceeds the pass budget it keyword-ranks segments (+neighbours for context) and flags the answer as partial rather than silently truncating. New meeting_qa model role (tier-balanced default). - AskPanel: a Q&A panel in the meeting detail; each answer's segment citations are clickable and scroll the transcript to that moment (with a highlight flash) + cue the audio — provenance you can touch (§5.3). 4 new unit tests for the keyword/selection logic (incl. the match-wins- budget-over-its-own-context case). Full suite 1651 passed; all 18 /notes routes register; frontend tsc clean. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016xdbRumGveUBFRsPh3Duu8 * refactor(notes): route STT through the LiteLLM plumbing + tier-stt model config STT was calling Groq/Deepgram/OpenAI via its own httpx clients with its own key-preference resolution — bypassing the platform's model plumbing that every other app uses. This makes speech-to-text a configured model like everything else: - New tier-stt tier: added to infra/litellm/config.yaml (default groq/whisper-large-v3-turbo), acb_llm _TIER_ALIAS_MAP/_TIER_DEFAULTS, and settings _TIER_LABELS — so it resolves through the same _TIER_MODEL machinery as the chat tiers and is editable in Settings → Models (the frontend already declared an 'stt' tier). set_tier_model now accepts it. - acb_stt.LiteLLMSTT: resolves the model via resolve_underlying_model( 'tier-stt'), loads keys with _ensure_keys_loaded, registers via ensure_model_registered, transcribes with litellm.atranscription (verbose_json → segments/words), and emits usage via _emit_usage for observability — mirroring acompletion_with_fallback. Removed the bespoke httpx providers + key-preference registry. NOTES_STT_PROVIDER/ DEEPGRAM_API_KEY env replaced by an optional NOTES_STT_MODEL override. - pipeline: transcript_source is now the resolved litellm id (e.g. groq/whisper-large-v3-turbo), no double prefix. Caveat: diarization is provider-dependent (whisper-via-litellm returns none → capture-channel prior); real diarization arrives with the self-host tier (slice 3), which becomes just another tier-stt model. Verified: tier-stt resolves to the configured model via the shared tier init; 8 rewritten acb_stt tests (litellm response normalization + tier resolution) pass; full suite 1652 passed; gateway imports clean; no blocking lint. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016xdbRumGveUBFRsPh3Duu8 * feat(notes): scratch-notes merge — jot during a meeting, AI expands it (Granola pattern) The single highest-leverage note-taker UX idea (research §3, spec §4 Tier-1 item 3): the user's attention during the meeting steers a far better summary than the transcript alone. - migration 97_note_taker_scratch.sql: meeting.scratch_notes TEXT (the user's rough input notes, distinct from the generated meeting_note output). Validated on a clean 00->97 chain + idempotency. - summaries.py: generation reads scratch_notes and threads it into the prompt via _scratch_block as EMPHASIS SIGNALS — expand the flagged topics from the transcript, fix shorthand, ensure each is addressed, but ground every fact in the transcript and never invent from the notes. Works for both single-pass and map-reduce. - meetings.py: PUT /meetings/{id}/scratch; scratch_notes surfaces on the meeting detail. - UI: a 'Jot notes as you go' textarea on the recording screen (debounced autosave, flushed on stop) and a 'Your notes' editor on the meeting detail (save on blur) — both feed the same field. Rebased onto current main (picked up #189/#190). 3 new tests for the scratch block (empty no-op, labelling + grounding guardrail + bound); full suite 1694 passed; 19 /notes routes; frontend tsc clean. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016xdbRumGveUBFRsPh3Duu8 * feat(notes): org glossary — teach transcription your jargon (§4 Tier-1 item 6) A per-user vocabulary that biases transcription toward the right spellings of product/people/customer names and acronyms — fixing the most common source of transcript errors before it propagates into notes, action items, and search. - migration 98_note_taker_glossary.sql: notes_glossary(user_id, term) with case-folded per-user uniqueness. Validated on a clean 00->98 chain + idempotency; ORM model added. - glossary.py: GET/POST/DELETE /notes/glossary CRUD (add is idempotent per case-folded term) + format_glossary_prompt/glossary_prompt — a bounded vocabulary-hint sentence (term/char caps) fed to the STT prompt. - pipeline.py: loads the meeting owner's glossary and passes it as SttOptions(prompt=...) at transcription time; best-effort so it never blocks transcription. - UI: a GlossaryModal (chip add/remove) opened from a Glossary button in the /notes library header. 4 new tests for the prompt builder (empty no-op, vocabulary-hint framing, trimming, length bound). Full suite 1698 passed; 21 /notes routes; frontend tsc clean. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016xdbRumGveUBFRsPh3Duu8 * feat(notes): selectable STT model + Deepgram named speakers (Settings → Models) Make the Note Taker's speech-to-text model user-selectable and add named speakers (diarization) via a paid Deepgram model — all routed through the existing LiteLLM tier plumbing. Self-hosted STT (slice 3) is deferred. Settings → Models - The Tiers tab gains a dedicated "Speech-to-text" section: its own editable picker for the `tier-stt` model, filtered to transcription-capable models (chat pickers exclude them, and vice versa). STT is no longer a "Coming soon" card. Fixes the tier→metadata lookup (tier-fast → fast) so labels/icons render. - New `ModelInfo.transcription` capability flag drives the split; Deepgram gets a provider tile, setup guide, colour/icon, and a "names speakers" badge. Model catalogue (settings.py + acb_llm.model_limits) - Add whisper models to Groq/OpenAI, gpt-4o-transcribe, and a Deepgram provider (nova-3/nova-2, DEEPGRAM_API_KEY). All tagged transcription:True, audio:True. - `_is_transcription_model` heuristic flags un-catalogued whisper/nova variants a live provider fetch surfaces, keeping them out of the chat pickers. - `_provider_from_model` recognises deepgram/*; key_store loads DEEPGRAM_API_KEY. Diarization plumbing (acb_stt) - `transcribe()` branches params per family: Deepgram gets native diarize/punctuate/smart_format; whisper keeps verbose_json + glossary prompt. - `normalize_transcription` reconstructs speaker-attributed segments from Deepgram's word-level speakers (litellm strips them from `words` but leaves the raw JSON on `_hidden_params`); whisper's segment path is unchanged. - `capabilities()` reports diarization based on the resolved model. Verification: 3 new acb_stt tests (Deepgram hidden-params diarization, param branching); full unit suite 1700 passed; blocking ruff clean; frontend tsc + next build clean. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016xdbRumGveUBFRsPh3Duu8 * fix(notes): consolidate slice-2 migrations into 99_note_taker_slice2 Rebased onto current main. main now carries 96_gtd_item_deep_work, 97_gtd_planning_prefs and 98_gtd_day_templates (the last two from the #198 collision fix), which collide by number with this branch's 96_note_taker_attendees / 97_note_taker_scratch / 98_note_taker_glossary. Both the migration runner (apply_migrations.sh) and test_migration_prefixes only recognise TWO-digit prefixes ([0-9][0-9]_), so 99 is the only free slot and the three can't each take one. They're additive, idempotent, and order-independent (each depends only on 95_note_taker or base uuid support), so they collapse cleanly into a single 99_note_taker_slice2.sql. Spec + ORM comments repointed to the new filename. Verification: test_migration_prefixes + test_acb_stt pass; full unit suite 1710 passed; blocking ruff clean. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016xdbRumGveUBFRsPh3Duu8 --------- Co-authored-by: Claude <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why
Agents can create interactive generative UI, but they weren't reaching for it — and one whole agent type couldn't be reached by the guidance at all. This closes both gaps. The review confirmed the emit path and frontend render are fully runtime-agnostic (one shared
queue.put({CUSTOM, generative_ui})viaresolve_run_queue(session_id), rendered on payload alone), so nothing was wrong with UI creation — the problem was entirely on the prompting side.Two root causes
emit_generative_uitool docstring (a weak signal the model consults after it decides to look at tools) and, at system-prompt level, buried mid-addendum among ~15 other tools. Nothing prompted the model, at answer time, to convert a data / status / comparison / choice reply into UI._build_injected_tools_addendum(which carries the genUI guidance anddesign.md) is applied only on the GitHub-Copilot_toolsbranch of_inject_agent_tools. Native MAF agents — e.g.email-assistant, which explicitly scopes inemit_generative_ui— got the tool + its docstring + at most the delegation registry. The agent type most set up to render UI had the weakest guidance.Fix —
_ui_first_directiveA static, byte-stable (KV-cache safe), tool-gated "Rich UI by default" directive delivered to both runtimes:
--cc-*CSS-token hook for when it's unavoidable). This directly answers "don't burn tokens on custom generation" and "use the proper design language, fonts, themes, icons."default_optionsandagent.toolsshapes — marker-guarded for idempotency, with a compact variant for sub-agents. MAF and Copilot agents now get the identical answer-time rule.formCard/optionPicker+"hitl":true.Design-language scope note: the full 16 KB
design.mdis intentionally not dumped into every MAF prompt — templates are on-brand without it, and that's the path the directive steers to. Injectingdesign.mdfor MAF agents that lean on custom-HTML reports is left as an optional follow-up (documented in the spec); the--cc-*hook covers the common custom-HTML case.Tests
tests/unit/test_genui_proactive_directive.py(7 tests): native-MAF instructions injection for both MAF shapes, idempotency (marker guard), compact-vs-full variant selection, byte-stability, template-first ordering.evals/trajectories/test_tool_addendum_drift_trajectory.py: new assertions that the directive leads both addenda, templates precede the custom-HTML escape hatch, and the directive is gated on the tool being present. Also refreshed the drift set to cover the coding-skill tools (run_script/code_task/list_integrations) that were injected but untracked.ai-company-brain/specs/generative_ui_2.md§7.🤖 Generated with Claude Code
https://claude.ai/code/session_01K9rzBFbJiEtwSyq8u7sWWt
Generated by Claude Code