Changelog

All notable changes to this tile are documented here.

0.1.28 — 2026-07-21

CI — PR-time fleet-review trigger

Added a thin review-trigger.yml that dispatches an immediate single-PR review in jbaruch/coding-policy so the policy verdict lands before merge; the coding-policy cron poll stays the backstop. No tile-content change.

0.1.27 — 2026-07-21

CI — Move to the central fleet policy reviewer

Migrated from the per-repo review workflow to the central jbaruch/coding-policy fleet reviewer (one ChatGPT-subscription credential held only in coding-policy). No tile-content change.

0.1.26 — 2026-07-20

CI — Repoint copilot-instructions at the review workflow

Docs-only pointer update accompanying the reviewer migration. No tile-content change.

0.1.25 — 2026-07-20

CI — Migrate to the Codex CLI subscription reviewer

Switched the policy reviewer to the Codex CLI authenticated by a ChatGPT subscription (no API key). No tile-content change.

0.1.24 — 2026-07-19

Fix — check-orders catches Shopify orders via the transactional sending subdomain (jbaruch/nanoclaw-orders#44)

The Shopify query from:noreply@shopify.com OR from:no-reply@shopify.com matched nothing — Shopify never sends order mail from shopify.com. A live sample found Shopify order mail splits two ways: platform-hosted senders on store+NNN@t.shopifyemail.com (55 threads, all transactional) and merchant-custom-domain senders (pacagen.com, knifeaid.com, talesofvalhalla.com) with no queryable Shopify signal. The existing subject-keyword queries recalled only 28 of the 55 platform-hosted orders (query 5 caught zero — Shopify subjects read Order #NNN confirmed, never the order confirmation bigram), so 27 Shopify orders were missed entirely.

Replace the dead query with from:t.shopifyemail.com — Shopify's transactional subdomain, ~100% order precision. Marketing rides the m./g. subdomains, so the swap brings in no promo bleed. Store-hosted senders on their own domain still depend on the subject/"Your order" queries — this recovers the platform-hosted stream, not every Shopify order.

Fetching the emails is not enough on its own: the recovered messages come from store+NNN@t.shopifyemail.com with subjects like Order #NNN confirmed, which the old prose maps would have classified as source="other" / status="unknown". The source/status classification moves out of the Step 4 prose table into a new tested scripts/classify-order.py: shopifyemail.com (any subdomain) normalizes to shopify, and confirmed joins the ordered keywords, with status priority ordered so a shipment/cancellation/refund signal still outranks a bare confirmed. Merchant-custom-domain senders remain source="other" (no queryable Shopify signal) — the documented limitation.

Surface sync: skills/check-orders/scripts/fetch-order-emails.py (the QUERIES constant + rationale comment), skills/check-orders/scripts/classify-order.py (new), skills/check-orders/SKILL.md (Step 4 source/status now call the classifier), tests/test_classify_order.py (new), README.md (script list). The query and map logic live in the scripts per coding-policy: script-as-black-box.

0.1.23 — 2026-07-18

Fix — check-orders reads the order total from the email body (jbaruch/nanoclaw-orders#38)

Step 4's amount rule read $XX.XX from subject+snippet only and took the largest match. A live sample of real order emails (the 5 production queries) found the total sits ONLY in the body for 8 of 9 order confirmations — Amazon Ordered:/Shipped: and store-hosted Shopify confirmations print nothing but Ordered: <item> / Order #X confirmed above the ~200-char snippet fold — so the old rule defaulted amount to 0 for essentially every order. The fetched-but-unread body field (#37 removed a contract sentence that claimed Step 4 read it) is now read by a named rule, resolving the "read or don't project" question in favour of read.

New scripts/extract-amount.py owns the extraction: prefer a labeled Order Total / Grand Total / Total line (searched subject→snippet→body, last match wins) over any largest-amount pick, and restrict the unlabeled largest-amount fallback to subject+snippet. The sample showed why "largest in body" is wrong — two of three Shopify bodies carried a struck-through list price ($98.99, $69.00) larger than the true total ($94.14, $62.10); a labeled total is never a struck price. The subtotal, sub-total, and sub total spellings are excluded from the bare-Total match. tests/test_extract_amount.py covers the body-only total, the struck-price trap, refund-in-snippet, and subtotal disambiguation.

Not addressed here (separate follow-up): the from:*@shopify.com production query matches nothing, because Shopify sends order mail from each merchant's own domain — real Shopify confirmations are caught only by the subject/"Your order" queries.

Surface sync: skills/check-orders/scripts/extract-amount.py (new), tests/test_extract_amount.py (new), skills/check-orders/SKILL.md (Step 2 body note, Step 4 amount rule + invocation), skills/check-orders/scripts/fetch-order-emails.py (docstring + projection comment), README.md (script list).

0.1.21 — 2026-07-18

Fix — nightly-order-sync cadence cap drops below the cron-interval multiple (jbaruch/nanoclaw#803)

precheck-nightly-order-sync.py set CADENCE = timedelta(days=3) — an exact multiple of the 24h daily cron interval. The cursor stamps at run completion, so the intended every-third-day fire lands ~71.8h old, age >= CADENCE fails, and the run slips to every fourth day (live task_run_logs: ran 07-11, then 07-14). This is the fleet-wide near-miss jbaruch/nanoclaw-admin#353 / jbaruch/nanoclaw-admin#354 first fixed elsewhere. The cap drops to 60h: above the 48h two-day fire so the run isn't every other day, below the 72h near-miss with a half-period of slack that also absorbs the local-TZ cron's DST drift. The three_day_boundary test becomes a three_day_near_miss regression guard; the cap value is de-hardcoded from SKILL.md, state-schema.md, and references/cadence-rationale.md per coding-policy: script-as-black-box.

0.1.20 — 2026-07-16

Changed — migrate the order-email fetch from Composio to native Gmail REST (jbaruch/nanoclaw#638)

fetch-order-emails.py loaded composio-rest.py from nanoclaw-admin's heartbeat skill over the co-loaded tessl__heartbeat mount. #638 deletes that file: Google access now goes to the native Gmail REST API through the OneCLI TLS-MITM gateway, which owns the OAuth connection and injects Authorization: Bearer on the wire. Separate registries mean admin's publish breaks this tile's 06:15 nightly fetch until this one ships, so this lands immediately after admin.

The fetch is rewritten onto heartbeat's new foundation — google-rest.py (transport), gmail-ops.py (list/get), gmail-message.py (native message parsing) — and sanitize-email-body.py now exports only sanitize(), its Composio-shaped sanitize_message() / DEFAULT_FIELDS having died with the invented field names they mapped. The stdout contract is unchanged, so SKILL.md Steps 3-10 are untouched.

Two shape facts drove the rewrite. Native users.messages.list answers a query with {id, threadId} stubs only, where Composio's GMAIL_FETCH_EMAILS returned full messages — so every message costs a second get, and dedup now runs against the stubs before any body is paid for (an id surfaced by three of the five queries costs one get, not three). And native snippet is a real ~200-char preview rather than Composio's {body, subject} object, so snippet and body map to genuinely distinct fields instead of needing the _as_text() coercion — that helper, the envelope-shape sniffing, the successful: false branch, and the messageId or id / from or sender dual reads were all Composio artifacts and are deleted rather than ported. internalDate (epoch milliseconds) replaces Composio's date field.

Credential handling is gone, not relocated: no COMPOSIO_API_KEY / COMPOSIO_USER_ID, no preflight, nothing in the container's environment to miss. The two failure modes it hid are actionable instead — a 401 (GatewayNotInjecting: the gateway is off the request path or the app is disconnected) and a 403 access_restricted (TierAccessRestricted: the untrusted tier is gated from Google by design) both exit 2 with a remediation rather than being retried as if transient. Per-query error isolation is preserved and extended: an unreadable individual message is an order alert that will not fire, so it becomes an errors[] entry against the query that surfaced it rather than being dropped silently. Step 2's error table reads on "all 5 queries errored and nothing was fetched", so a partially-failing run still advances the cursor — the forward-progress property references/write-ahead-rationale.md exists to protect.

Surface sync: skills/check-orders/scripts/fetch-order-emails.py, skills/check-orders/SKILL.md (Core Rule, Step 2, error table), skills/check-orders/references/write-ahead-rationale.md, README.md (capabilities, tier rationale, required environment, cross-tile dependency, script list), tests/test_fetch_order_emails.py, tests/conftest.py, and tests/fakes/ — composio-rest.py deleted, sanitize-email-body.py trimmed to sanitize(), plus new native-shaped doubles for google-rest.py, gmail-ops.py, and gmail-message.py.

0.1.19 — 2026-07-13

Changed — align nightly-order-sync wording with the sqlite order store (jbaruch/nanoclaw-orders#32)

The nightly-order-sync skill still called the order store orders-db in its frontmatter description and Step 1. That literal orders-db.json file was migrated into the sqlite orders table in #294 (2026-04-30); the canonical name is now the orders table in /workspace/store/messages.db, per check-orders/SKILL.md. Replace the stale shorthand in both spots. Documentation wording only — no behavior change.

0.1.18 — 2026-07-08

Changed — bump github/gh-aw-actions/setup to v0.82.5 (jbaruch/nanoclaw-orders#30)

Dependabot retargeted its jbaruch/nanoclaw-orders#9 from 0.82.2 to 0.82.5 minutes after jbaruch/nanoclaw-orders#29 landed 0.82.2 — the registry moved mid-campaign. Bump the pin in both compiled gh-aw review locks (uses lines, comment inventory, and embedded gh-aw-manifest entries) to v0.82.5. No workflow logic change.

0.1.17 — 2026-07-08

Changed — refresh pinned GitHub Actions (jbaruch/nanoclaw-orders#29)

Adopt the four open Dependabot GitHub Actions bumps as one CI-scoped change, with each bump kept as its own Dependabot-authored commit: actions/checkout v4 → v7 (test.yml, publish-tile.yml — also silences the runners' Node 20 deprecation warning), actions/setup-python v5 → v6 (test.yml), plus actions/cache/restore v5.0.5 → v6.1.0 and github/gh-aw-actions/setup v0.81.6 → v0.82.2 in the compiled gh-aw review locks (Dependabot maintains those pins between gh aw compile runs). Review follow-ups: actions/cache/save is bumped to the same v6.1.0 Dependabot left it trailing at v5.0.5, and the locks' embedded gh-aw-manifest entries are updated to match the visible pins. Supersedes jbaruch/nanoclaw-orders#6–#9, which lacked the CHANGELOG blocks the stamp step needs. No workflow logic change.

0.1.16 — 2026-07-08

Changed — bump pytest to 9.1.1 (jbaruch/nanoclaw-orders#28)

Dev-toolchain pin refresh across a pytest major version. The suite runs clean at 9.1.1 — 130 passed, zero skips, no deprecation warnings — no code change. Supersedes Dependabot's jbaruch/nanoclaw-orders#11, which lacked the CHANGELOG block the stamp step needs.

0.1.15 — 2026-07-08

Changed — bump pyright to 1.1.411 (jbaruch/nanoclaw-orders#27)

Dev-toolchain pin refresh. Supersedes Dependabot's jbaruch/nanoclaw-orders#12, which lacked the CHANGELOG block the stamp step needs. Zero findings at the new version — no code change.

0.1.14 — 2026-07-07

Changed — freeze wall-clock tests and unskip fail-closed coverage (jbaruch/nanoclaw-orders#20, jbaruch/nanoclaw-orders#21)

Two test-hygiene fixes. First (#21): test_flag_anomalies.py and test_promote_stale_shipped.py computed fixtures from date.today() / datetime.now(), so the boundaries under test moved every day; both now freeze module.date / module.datetime to fixed test doubles (the test_within_days.py pattern) with literal fixture dates. Second (#20): the fetch-order-emails.py fail-closed main-path tests were skipped as heartbeat-dependent; new local fakes under tests/fakes/ mirror the heartbeat modules' load surface, unskipping the missing-credentials and missing-REST-helper paths, and the two sanitizer-behavior skips are replaced by a contract test proving the loaded sanitizer runs before projection. The suite now runs with zero skips.

0.1.13 — 2026-07-07

Changed — move Step 6 exclusion matching into deterministic code (jbaruch/nanoclaw-orders#18)

Step 6 specified address parsing, multi-recipient handling, case-insensitive comparison, and the description fallback as prose for the agent to re-implement on every run — a mismatch could re-flag orders the user explicitly excluded. New apply-exclusions.py owns the EXCLUSIONS table and all matching (recipient parsing via email.utils.getaddresses), unflags matches in one transaction, and emits the id list Step 8 passes as EXCLUDED_IDS. Tests cover display-name wrapping, comma-separated recipients, case-insensitivity, the NULL to_address description fallback, and same-domain non-matches. unflag-orders.py stays as the ad-hoc unflagging utility.

0.1.12 — 2026-07-07

Fixed — HTML-escape untrusted fields in Telegram order alerts (jbaruch/nanoclaw-orders#19)

Step 10 interpolated raw description / flag_reason / source / order_date into Telegram HTML. description derives from email subject text — untrusted, sender-controlled — so a subject carrying <, >, &, or tags could break the message parse (suppressing the alert entirely) or inject links/formatting into the notification. New render-order-alerts.py owns the rendering: it HTML-escapes every field and emits {"message": <str|null>, "count": <int>} with the ready-to-send text; the SKILL pipes get-flagged-orders.py into it and sends message verbatim (null → stay silent). Hostile-input tests cover A&B <tag> and </b><a href="...">x</a> descriptions plus non-object array elements.

0.1.11 — 2026-07-07

Fixed — backfill missing release entries and guard the version/CHANGELOG sync (jbaruch/nanoclaw-orders#17)

Versions 0.1.7–0.1.9 published without CHANGELOG entries: their PRs added no un-headed entry blocks, so the stamp step (wired in 0.1.7 itself) had nothing to stamp and tile.json advanced while the CHANGELOG stood still. Backfill the three sections from the merge commits that produced each release, and add tests/test_changelog_sync.py — a pytest guard asserting the first ## X.Y.Z heading in CHANGELOG.md matches tile.json's version, so the next entry-less release fails the very next PR's CI instead of drifting silently for three versions.

0.1.10 — 2026-07-07

Changed — bump ruff to 0.15.20 and reformat (jbaruch/nanoclaw-orders#16)

Land the ruff 0.7.4 → 0.15.20 toolchain bump together with the mechanical reformat it forces on three files (implicit string-concat collapse, assert-message wrapping). Under the previously pinned 0.7.4 the tree was format-clean — the drift issue #16 reported reproduces only under newer ruff, which is exactly what was blocking Dependabot's ruff bump PR #10 on the format gate. Bump and reformat land as one PR; either half alone leaves CI red. Supersedes PR #10.

0.1.9 — 2026-07-07

Changed — ignore tessl-generated .github/mcp.json (jbaruch/nanoclaw-orders#15)

Current tessl CLI init emits .github/mcp.json (GitHub Copilot CLI MCP config) alongside the scaffolding the existing ignore block already covers; the block predated it and let it leak through as untracked noise. Entry backfilled by #17 — the release shipped without one.

0.1.8 — 2026-07-03

Changed — refresh coding-policy PR review workflows

Upgrade the gh-aw jbaruch/coding-policy PR review workflow templates (OpenAI and Anthropic reviewers, sources and compiled .lock.yml forms) to the latest published version. Entry backfilled by #17 — the release shipped without one.

0.1.7 — 2026-07-02

Changed — wire coding-policy stamp-changelog step before publish (jbaruch/nanoclaw-orders#14)

Add the jbaruch/coding-policy stamp-changelog action immediately before tesslio/patch-version-publish, matching nanoclaw-travel: authors add un-headed ### CHANGELOG blocks and the step writes the ## <version> — <date> heading at publish time. Entry backfilled by #17 — the release that wired the stamper shipped, fittingly, without one.

0.1.6 — 2026-07-02

Changed — backfill CHANGELOG entries for released versions 0.1.1–0.1.5

Versions 0.1.1, 0.1.3, and 0.1.4 shipped without CHANGELOG entries, and the 0.1.2 agentModel note sat un-versioned at the top of this file. Every released version now has a heading; the entries are reconstructed from the merge commits that produced each release. No code change.

0.1.5 — 2026-07-02

Added — gate language diagnostics in CI with pyright (jbaruch/nanoclaw-orders#2)

Adopt a pyright zero-findings gate: pyrightconfig.json for the skill-bundle layout and a python -m pyright --warnings skills/ tests/ CI step after ruff, before pytest (--warnings fails on warnings too). The first run surfaced two real bugs — stamp-cursor.py built its argparse description from __doc__.splitlines()[0], which crashes at startup under python -OO (docstrings stripped to None), and the test module loaders subscripted a ModuleSpec | None. Fixed with a literal description and explicit if ...: raise guards, no suppressions. Adds a weekly Dependabot for the pinned dev toolchain.

0.1.4 — 2026-07-02

Changed — refresh coding-policy PR review workflows (jbaruch/nanoclaw-orders#4)

Upgrade the gh-aw jbaruch/coding-policy PR review workflow templates to the latest published version.

0.1.3 — 2026-07-01

Changed — refresh coding-policy PR review workflows (jbaruch/nanoclaw-orders#3)

Upgrade the gh-aw jbaruch/coding-policy PR review workflow templates to the latest published version.

0.1.2 — 2026-06-08

Changed — pin nightly-order-sync to Haiku via agentModel: (jbaruch/nanoclaw#613)

Order-email data sync is triage, not synthesis. Pin agentModel: "claude-haiku-4-5-20251001" in the skill's frontmatter so it stops defaulting to Opus (cadence-registry plumbs it to scheduled_tasks.agent_model). Full model ID matches the gateway model_list row. Part of the #613 Claude tier-down.

0.1.1 — 2026-06-07

Added — script tests omitted from the initial scaffold

Add the script unit tests that were left out of the initial tile scaffold.

0.1.0

Added

• Initial tile: check-orders + its nightly-order-sync cadence wrapper migrated from nanoclaw-admin into a standalone public per-chat overlay tile (jbaruch/nanoclaw-admin#319). check-orders fetches order-related Gmail over the Composio v3 REST endpoint, upserts the orders table in the orchestrator's messages.db, and flags cancellations/refunds, large purchases until delivered, and overdue deliveries — staying silent on normal order flow. nightly-order-sync runs it on a 3-day-capped 15 6 cadence and emits the observable-silence cursor marker the silent-success watchdog reads. Co-locating the cadence driver with the skill it drives keeps the orders domain self-contained in one tile (same pattern as nanoclaw-conferences's check-cfps + nightly-cfp-sync) and removes the cross-tile Skill() call that would otherwise span admin → orders. Carries the ten check-orders helper scripts, the two nightly-order-sync scripts, and both clusters' tests unchanged from the admin originals.
  • Origin: nightly-order-sync was peeled off the nightly-external-sync bundle in jbaruch/nanoclaw#581 so order fetching gets its own bounded container instead of being cut off in the bundle's long tail.

Cross-tile dependencies

• skills/check-orders/scripts/fetch-order-emails.py loads sanitize-email-body.py and composio-rest.py at runtime from nanoclaw-admin's heartbeat skill via the co-loaded tessl__heartbeat tile mount; the orders state (orders + orders_metadata tables) lives in the orchestrator's shared messages.db (read-write on main/trusted), where admin's morning-brief and check-email read flagged orders. These resolve because admin co-loads with this overlay in the owner's main/trusted chat. The Gmail fetch path still rides on Composio, which jbaruch/nanoclaw#637–#640/#564 are retiring in favor of OneCLI — the moved fetch script will be reworked when that lands. Tests that assert the heartbeat sanitizer's own behavior (body cap, invisible-unicode collapse) or load the real heartbeat files are skipped here with a tessl__heartbeat-dependency reason; that coverage lives in admin's heartbeat suite. The orders-owned projection, dedup, field-fallback, cursor, and fail-closed-on-missing-sanitizer logic stays fully tested with an identity sanitizer double.

Rules

• Closed-loop carve-out claimed for jbaruch/coding-policy: plugin-evals (2026-06-07). 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. Covers both decisional skills in this tile (check-orders, nightly-order-sync). No evals/ directory ships in this tile.