Changelog

All notable changes to this tile are documented here.

0.1.8 — 2026-07-01

Changed

• Bumped the pinned ruff from 0.7.4 to 0.15.20 and reformatted the tree to match. Pure style (0.15.20 collapses adjacent implicit string concatenations that now fit on one line); no behavior change. Doing the bump-plus-reformat together keeps a bare Dependabot version-bump from landing red against the ruff format --check gate.

0.1.7 — 2026-06-30

Changed

• The check-cfps Sessionize round-trips moved out of the agent's context into deterministic scripts that call the Sessionize universal API directly, closing the "skip to save tokens" hole behind issues jbaruch/nanoclaw-conferences#4/#7/#8/#9. The MCP sessionize_* tools stay available for ad-hoc queries; only the pipeline path changed. Because a Python subprocess can't invoke an MCP tool, the ~260 KB open-CFP and ~267 KB events payloads used to land in the agent's own context, where a token-pressured Haiku run would improvise inline Python — misparsing the open-CFP list as "1 event" (0 new candidates on 2026-06-29) and skipping the live verify call while faking verdicts from memory yet recording success.
  • Verification driver (#7): new scripts/verify-sessionize.py collapses Step 5's prepare → events round-trip → apply into one invocation. It fetches each slug from https://sessionize.com/api/universal/event?slug= (X-API-KEY: SESSIONIZE_EVENT_API_KEY), normalizing each raw event in-tile via a direct port of the host's normalizeSessionizeEvent. Per-slug failures isolate to verify_failed (bounded concurrency, the host's batch contract) and never substitute remembered verdicts.
  • Discovery (#9): new scripts/discover-open-cfps.py calls .../open-cfps (X-API-KEY: SESSIONIZE_SPEAKER_KEY), applies the host's online/user-group filter, and emits the Step 2 candidate count deterministically instead of parsing the payload inline.
  • Evidence-gated freshness (#8): scripts/stamp-last-checked.py no longer advances _last_checked unconditionally. The driver writes a verify-evidence.json marker, and the stamp advances the heartbeat only when ≥1 entry was resolved from a live response this run (or there was nothing to verify). A skipped call or total Sessionize outage now records _last_checked_skipped and exits 3 — so the run can't report a clean success or fool the jbaruch/nanoclaw#601 work-evidence watchdog. nightly-cfp-sync reads the new verification field and, on "skipped", holds the cadence cursor and retries next fire instead of resting 72h on an unverified run.
  • The run-state checkpoint store's prep/sessionize_results/decisions stages collapse into a single verify stage.

Required environment

• The CFP pipeline now reads two host-injected Sessionize keys from the container environment: SESSIONIZE_SPEAKER_KEY (open-CFP discovery) and SESSIONIZE_EVENT_API_KEY (event verification) — the same keys the NanoClaw host already held for its sessionize_* tools. SESSIONIZE_API_BASE optionally overrides the API base. See .env.example and README "Required environment."

0.1.6 — 2026-06-30

Added

• Pyright is now gated in CI at zero findings (jbaruch/nanoclaw-conferences#6, adopting jbaruch/coding-policy: language-diagnostics): new pyrightconfig.json (basic, py3.11, over skills + tests), a pinned pyright==1.1.390, and a python -m pyright step between ruff and pytest. Turning the gate on for a never-checked tree is landed as its own focused change, separate from feature work, per the rule.
• .github/dependabot.yml gives the pinned dev dependencies (pytest, ruff, pyright) and the GitHub Actions a stated, automated renewal mechanism, per jbaruch/coding-policy: dependency-management.

Fixed

• The first pyright run surfaced 13 real findings, all fixed with typed None-guards (no blanket suppressions): unchecked importlib spec_from_file_location (ModuleSpec | None) access in the prepare-sessionize-batch.py sibling loader, the conftest.py loader, and three test-module loaders; and __doc__-is-None (under -OO) before .splitlines()[0] in stamp-cursor.py.

0.1.5

Added

• check-cfps now stamps an honest freshness heartbeat and resumes interrupted runs from disk (jbaruch/nanoclaw-conferences#4).
  • Freshness: new deterministic scripts/stamp-last-checked.py is the single writer of cfp-state.json's top-level _last_checked, run in Step 8 on every successful pass. It stamps the run timestamp unconditionally (a run that re-verified everything and changed nothing still "checked"). Replaces LLM hand-writing, which had left the field frozen at 2026-04-22 while every record had refreshed days earlier — on 2026-06-13 that stale field drove a wrong "pipeline stalled ~7 weeks" live diagnosis when the pipeline was in fact healthy (running on its ~72h cadence). _last_checked is distinct from the wrapper's nightly-cfp-sync-cursor.json last_run (which gates cadence); both exist on purpose.
  • Resumability: new scripts/run-state.py checkpoint store persists each pipeline stage artifact (fetch, candidates, prep, sessionize_results, decisions, working_set) under /workspace/group/state/cfp-run/, with a Step 1 resume guard and clear-on-success teardown in the SKILL. Motivated by the 2026-06-10 run, a token-limit continuation that had hand-rolled the whole pipeline inline, blew its budget mid-run, and — finding its prep output was never persisted — re-derived it and re-discovered the cfp-state.json schema from a chat summary. The store is per-UTC-day (a continuation on a later day resets to a fresh run); resume is best-effort since Step 5 re-verifies the full cohort, so a fresh full run is always safe. It is a flat directory of JSON files rather than a messages.db table — a short-lived, single-installation, single-writer scratch artifact with no cross-row queries, torn down on success — same basis as the nightly-cfp-sync cursor.

0.1.4

Changed

• Pinned nightly-cfp-sync to Haiku (claude-haiku-4-5-20251001) via agentModel: frontmatter — CFP data re-verify/sync is triage, not synthesis, so it no longer defaults to Opus. Part of the #613 Claude tier-down (jbaruch/nanoclaw#613).

0.1.3

Changed

• Reworded skills/check-cfps/references/web-fetch-fallback.md to describe fetch_markdown as the NanoClaw host's server-side renderer in neutral terms, dropping the snitchmd / CloakBrowser / "past anti-bot gates" wording the registry's intent-review moderation repeatedly misread as an attacker-proxy prompt-injection (false positive — these are first-party NanoClaw rendering services). Behavior is unchanged: same fetch_markdown-then-Composio fallback chain. Also dropped the cross-tile reference to the admin-only max-effort skill so the public tile is self-contained. Ends the per-publish moderation-override treadmill for this tile.

0.1.1

Added

• nightly-cfp-sync cadence wrapper migrated from nanoclaw-admin (jbaruch/nanoclaw-admin#298). It runs check-cfps on a 3-day-capped 30 6 cadence, consumes the CFP list internally, surfaces only a stale-verification notice, and emits the observable-silence cursor marker. Co-locating the cadence driver with the skill it drives keeps the CFP domain self-contained in one tile (same pattern as nanoclaw-flight-assist's sync-tripit + check-travel-bookings) and removes the cross-tile Skill() call that would otherwise span admin → conferences. Carries its precheck + stamp-cursor scripts, cadence-rationale.md / state-schema.md references, and both unit tests unchanged from the admin original.
  • Origin: nightly-cfp-sync was peeled off the nightly-external-sync bundle in jbaruch/nanoclaw#581 so the heavy full-cohort CFP verification gets its own bounded container instead of being cut off in the bundle's long tail — which had left task_run_logs.result empty. (Moved here from the reference doc per coding-policy: context-writing-style — incidents live in the CHANGELOG, not auto-loaded context.)

Rules

• Closed-loop carve-out claimed for jbaruch/coding-policy: plugin-evals (2026-06-06). This tile is part of the jbaruch/nanoclaw-* plugin fleet — a fully-automated agent loop satisfying all three preconditions of the rule's "Narrow exception for closed-loop automated systems with no human eval-result consumption" clause: (1) no human reviews eval output for this tile in any form (no eval scores, no lift deltas, no scenario-by-scenario diffs, no regression alerts); (2) no automated gate consumes eval results (no evals.yml workflow, no publish-tile eval step, no downstream dashboard or paging route); (3) the owner accepts that re-introducing any consumption of eval results later — whether human review OR automated gating — requires re-introducing evals first under the standard requirement. Matches the carve-out claimed by jbaruch/nanoclaw-admin on 2026-05-09 and inherited by every jbaruch/nanoclaw-* tile thereafter (e.g. jbaruch/nanoclaw-flight-assist, 2026-05-18). Covers both decisional skills in this tile (check-cfps, nightly-cfp-sync). No evals/ directory ships in this tile.

0.1.0

Added

• Initial tile: check-cfps skill migrated from nanoclaw-admin into a standalone public per-chat overlay tile (jbaruch/nanoclaw-admin#298). The skill discovers open conference CFPs across multiple sources, applies source-aware Sessionize verification and tiered AI relevance routing, and maintains persistent per-CFP state in cfp-state.json with owner-side schema-version migration. Carries the eight deterministic helper scripts and their test suite unchanged from the admin original.

Changed

• check-cfps description no longer claims to "extend the tessl tile version" — this tile is the tile version now, so the self-referential phrasing was dropped while preserving the trigger phrases.