Repository navigation
The decision monitor: replace the deleted briefing product, tasks 026-033 - #25
Merged
Merged
Conversation
…anning Backlog task 026, the first task of docs/ESTATE_PLAN.md. Establishes three properties every later phase relies on, and nothing else: - Secrets have one home. The Anthropic key is read from the OS keychain through keyring when a model client is built, never at import and never from .env, with no environment fallback. `estate auth` is the only write path and never prints a value. - Model calls have one door and it is logged. src/llm/audit.py writes each request before it is sent and the usage or error after. The client returns usage, chooses its model by role (Sonnet 5 triage, Opus 5 synthesis), and passes effort through output_config on the 1.x SDK. - A leaked key cannot be committed. detect-secrets runs in pre-commit and in the CI security job, which could not fail before this. Also: the plan itself with its decisions recorded; backlog 019/020/023 superseded by the laptop plan, 018 rewritten for it, 015/016 rescoped into Phase 1d, 022 updated; CLAUDE.md's through-line paragraph moved from Questions to nodes; types-PyYAML added so make typecheck passes in a fresh venv; requirements recompiled (drops dev pins nothing declared or imported).
…n for the decision monitor Backlog task 027. 8,200 lines removed, none of them reachable from a console script, a migration, or a live module. Each candidate was adversarially reviewed by an agent told to find a live caller, a build step, a CI step, a packaging entry or a standing commitment; none was found, and three independent hunters then found the second tier this list had missed (historical drop migrations, unread settings fields, an unused palette, a dev helper in production code, four dependencies nothing imports). backlog/027-sweep.md records every deletion and the two overridden "keep" notes (deduplicator, coverage_probe) with reasons. Also: the estate console script and src/estate/ fold into insightweaver (auth is a subcommand; credentials live in src/config/); the REPL is gone; FeedMatcher only loads config/feeds/; requirements recompiled. docs/PLAN.md replaces docs/ESTATE_PLAN.md. The estate graph and its sensors are deferred with reasons; what remains is the decision monitor. Tasks 028 to 031 are queued from it. The reachability test the sweep built found fourteen modules unreachable from any command -- the whole ingest, routing and model stack -- which is the finding this plan exists to close; it lands with task 029.
…erministic routing to watches Backlog task 028 (tasks 015 rescoped). The repository has had no way to grow its corpus since the legacy write path closed on 2026-08-31; now `insightweaver ingest` builds an adapter for every feed in config/feeds/ (RSS included) and runs them with bounded concurrency through the one store path, so a feed added to config reaches the database by construction. Tier 1: a watch's trigger clauses compile through the entity matcher's word-boundary and shouted-term rules (AND across a clause's fields, OR within a field, any clause fires, first index recorded) into a `routes` table, idempotent per pair, no model call -- proved by a test that blocks the import of src.llm while routing runs. The regression gate plants twenty whole-word matches and a hundred look-alikes in a thousand observations and checks the routed rows by identity; with the boundary anchors removed, 61 route instead of 20. `route --dry-run` reports per watch, and reports the unrouted count by source and by near-duplicate cluster, bounded so a rebuild cannot hand the O(n^2) grouper the whole corpus. The task file records what the adversarial review found and what was deferred.
…recorded, nothing retried Backlog task 029 (task 016 rescoped). The single stochastic component: a ClaudeAdjudicator registered as prompt version claude-v1, its prompt, effort, caps and response schema fingerprinted and pinned by a test so a changed prompt cannot keep the old name. The schema is derived from the pydantic verdict model with the range keywords the grammar rejects removed; pydantic enforces the ranges on receipt. `insightweaver adjudicate` asks about every routed pair without an answer under this version, commits each answer as it is recorded, and records a refusal, a truncated reply or an invalid verdict as a failed row with its audit id and cost, never retried. A missing key or a request the API rejects aborts the run instead of becoming verdicts. `--dry-run` prints the exact text and never builds a client. The client is synchronous now, raises a typed failure that carries the audit id and usage, and takes the structured-output schema as a request parameter. Replay reads exactly the routed pairs, bounds by routed observations, and refuses an unrouted corpus. The reachability test the sweep built lands here: every module under src/ now has a caller. The task file records what the adversarial review found: three blockers fixed before commit, and what was recorded rather than changed.
…ement instead of deletion Backlog task 030 (tasks 017 and 021 cut to what has data behind it). A watch_beliefs table holds every belief value with its source and date -- 'file' when the watch file set it at sync, 'principal' when the operator wrote it with `watch believe` -- and refuses UPDATE and DELETE from the ORM and from a sqlite3 shell alike. watches.belief is the registration value and is never overwritten; a file that is behind the operator's own updates shows as a change with a source. There is no update rule: belief moves only when the operator moves it. `watch resolve` grades a watch once, by hand, with no unresolve and no path from evidence. sync retires a watch that leaves the file and restores one that returns; it never deletes a row with history on it. Liveness has one definition: routing links open, unexpired watches; adjudication and replay skip retired and resolved ones and keep expired ones, because expiry is routing's gate and a date filter in replay would make it depend on the day it ran. models.py became a package, one module per table family, mechanically. The task file records the review's findings and the fixes.
… instead of failed A connection error, a timeout, a 429 or a 5xx was recorded per pair as a verdict; one bad hour would have marked every pending pair failed with no answer given. ModelCallFailed.outcome now says answered, rejected or unavailable, and only an answered failure is recorded. Non-API exceptions are audited and re-raised. A max_tokens cut during thinking is named as such. replay --commit refuses --limit, since commit would delete every row outside the bound. The prompt fingerprint includes the model's own schema so a bound enforced on receipt moves the version.
…rministic to the byte, and run select_brief reads evidence, watch state, routing recency and source state into plain data, pure in (db, as_of, since, position); render prints it in terminal or markdown with every citation carrying its observation hash and prompt version. Every heading prints with a zero when empty. A briefs row per render sets the next default window to the last brief before today, so a second run the same morning reproduces the first. run chains ingest, route, adjudicate and brief and stops at the first failure. After review: staleness is tested on routes.routed_at, not the item's date; evidence for a watch that expired or was resolved in the window is shown with its state; a failed fetch is printed as a failed attempt, never as a fetch, through one record_attempt stamp; routed pairs no adjudication has answered are counted in QUIET; a 200 that is not a feed is an error in the RSS adapter; beliefs are read as of the brief's moment; delivery precedes the briefs row; a Position error reaches the header whole. Golden end-to-end test over the recorded Federal Register week, a reprinting feed, an empty feed and an unreachable one, with a scripted model client and a frozen clock. make golden regenerates it.
…to the private files and syncs them The operator answers, the skill writes, the loaders validate and watch sync is the only write path; it proposes no decision, belief, claim or trigger (invariant 6). Task 024's interview half is done this way; its review loop stays parked until there are resolved watches to review.
…rewritten from the tree, with a test that they name only what exists Every sentence written from the code as of 2026-09-23 and reviewed against it: what each command writes, what the audit log holds, what the observation hash contains, what sync does to a belief the operator moved, what a rebuild does to QUIET's clock. The five principles keep their headings; the third says it is not implemented. The README shows the golden brief whole. tests/test_docs.py fails when a document names a command, option, make target or repository path the tree lacks; tests/test_invariants.py pins that nothing in src sends, serves or listens and that only replay deletes evidence. Writing the guide found that nothing created the schema on a fresh database: make db-init (src/database/migrations/create_schema.py) does. The example watches file named a source that matches nothing; .env.example gained DATABASE_URL; pyproject's URLs point at the real repository.
Configured twice, once in geographic/states/virginia.json and once in domains/civic_engagement.json with a slightly different URL (no www.), which is why ensure_source's exact-URL dedup never caught it. Kept the civic_engagement.json copy: higher relevance_score and tags (legislation, general_assembly, watchdog) that match state-politics watches better.
A first real run against 72 configured sources buried the four sections worth reading daily under ~70 lines of mostly-successful source status. A source that answered, whatever it returned, is now folded into a summary count (configured/fetched/failed/never fetched). A source whose last attempt failed is still printed in full, every time, with no flag needed -- invariant 5 (silence distinguishable from breakage) is about failures being visible, not about every source having its own line, and the failed ones are exactly what the operator needs to act on. --verbose restores the original always-list-everything header. Golden files regenerated (make golden) and the README's embedded brief excerpt resynced to match. backlog/031-brief.md gets a dated addendum rather than an edit to its original ACCEPTANCE/LANDMINES, since those record what was actually reviewed at the time. Also fixed a pre-existing mypy gap in tests/brief/test_render.py (two dict-splat test helpers untyped enough to fail the pre-commit mypy hook, which unlike 'make typecheck' does check tests/) hit while touching the same file for the header change.
This was referenced Sep 24, 2026
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
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.
The decision monitor: backlog tasks 026 to 033, plus a first real run
Replaces the old briefing product (deleted 2026-08-31) with a decision monitor: pre-registered
watches, deterministic routing, one model call per candidate pair, and a brief that is a
deterministic view over the result.
docs/PLAN.mdis the plan; eachbacklog/0NN-*.mdcarriesthe task's acceptance, out-of-scope, landmines, the decisions taken while building it, and the
review findings applied against it.
make checkis green at every commit (770 tests on thelast one).
45c1c56(026)8e0eb1a(027)estateintoinsightweaver; drop the REPLa091a80(028)dc7909a(029)58c1e9e(030)b372b52(029 follow-up)0cac270(031)run; a golden end-to-end test with a frozen clock1dac7cc(032)/onboardskill: an interview that transcribes, never authorseebd83d(033)tests/test_docs.pyfails when a document names a command, option, make target or path that does not exist;make db-initfor a fresh database0df9f44cfe3d1e--verbosefor the full list), added after the first real run's header buried the sections worth reading daily under ~70 mostly-successful source linesValidated by a real run, not just fixtures
After 033, the branch was onboarded with real Position/watch data (via
/onboard) and runend to end:
ingestagainst 72 configured sources,route, a realadjudicatecall againstthe live API, and
brief. Routing and adjudication behaved as designed -- two items thatmatched a broad trigger were correctly judged not-evidence by the model, including one case
where a name match was accurate but the underlying claim wasn't. One adjudication's stated
rationale was found to misdescribe the text it read even though its verdict was probably still
right; recorded as a live finding, not fixed here, since it's model behavior rather than a bug
in this codebase.
Decisions a reviewer may want to overturn
HORIZON_DAYS); the plan gives no number.runchains ingest, route, adjudicate, brief; syncing watches is a separate step.answered sources into a count by default; the source row keeps no separate time of the last
success.
Not yet done
🤖 Generated with Claude Code