Conversation
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Implements the §5.3 total order over a subject's claims: later occurred_at wins, tie-broken by reliability, then credibility, then recorded_at, then claim_key lexicographically. resolve() applies the credibility floor (only claims at or above it are eligible to win) with an only-claim exception when a lone member is floored.
Implements the free, deterministic rules tier of the cascade classifier (spec §5.6): compile() pre-compiles a KindTaxonomy's regexes into a RuleSet, classify() returns the first matching kind in config order with confidence 1.0 and JudgeSource::Rule. A rule matches when any of its matchers match (any_of). Invalid rule regexes are rejected at compile() with ClogError::Corrupt, documented as a deliberate v1 fit (config-as-state) pending a possible dedicated InvalidConfig variant.
Adds the render module: a §5.8 grammar parser (SlotName, Segment,
Template, parse) that scans for %{...} slots and rejects unknown
names/keys, unterminated slots, empty slots, and non-usize limit
values as ClogError::TemplateError without ever panicking; the
byte-for-byte DEFAULT_TEMPLATE; and rfc3339_utc via Howard Hinnant's
civil-from-days algorithm. Renderer body itself lands in the next
task.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Implements the render/mod.rs slot renderer + budgeter (spec §5.8):
headline() normalization, the UrgentItem/LoopItem/EntityItem/ChangeItem/
SlotInputs engine-to-renderer contract, and render() itself, which
substitutes each %{...} slot's formatted item lines into the template and
budgets down to a char count by dropping whole items from the end of
slots in reverse priority order (changes, entities, open_loops, urgent),
appending/updating a "… (n more)" marker per truncated slot. Per-slot
limit= caps items before budgeting.
Corrected the u_tmpl_2 test's tight-budget derivation from full.len() - 1
(bytes) to full.chars().count() - 1: the render contains three U+00B7
middle dots, each 2 bytes but 1 char, so a byte-derived budget was never
tight enough to trigger truncation against the spec-correct char-counting
budgeter, which would have made the test's own assertions vacuous. The
assertions themselves are unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… views Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ed allow Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…AS-3, INV-6) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adds wal.rs: an append-only, length-prefixed, CRC32-checked frame log that is clog's source of truth (spec §6.3, INV-11). open_dir replays every existing frame in order; a torn or corrupt tail (short read, absurd length, CRC mismatch, or a CRC-valid record that still fails to postcard-decode) is never applied, quarantined to wal.corrupt, and truncated from the log so reopening afterwards is clean (R2). append writes one frame and fsyncs per FsyncPolicy.
…1/3/4/5/9) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… masked change detection Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…at consolidation Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adds tests/props.rs per Task 16 brief, with two rulings applied: - norm() strips the whole header line + changes-since-last-brief section (not just the rev marker), matching tests/api.rs's established INV-3 normalization. - P6's generated claims each carry a fixed entity (p:shared-ent) so believed winners are readable via select(View::EntityState), which is keyed by entity and would otherwise see none of these claims. All 6 properties pass at cases=64 and at PROPTEST_CASES=256 with no counterexamples. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Ten-claim, three-client, two-scope fixture (spec §11.4) exercising belief resolution's freshness-over-trust limitation and its healing on retraction, the entities slot dropping subject-less entities, and merge_entities re-keying entity grouping. Adds one auxiliary `fyi` classification rule beyond the brief's four so cash-and-collections' fyi=0.5 weight is actually exercised by a live claim; documented in task-17-report.md along with the hand/script-verified score tables for every checkpoint.
R1 re-executes the recovery test binary as a child that aborts in the test-crash hook immediately after each of the script's four WAL appends (post-fsync, pre-apply), then reopens that directory and asserts the recovered world is byte-identical to a fresh instance fed exactly the durable prefix -- same text, same scope rev, same as_of -- and that the survivor still writes, taking rev N+1. R3 pins the rebuild_on_open contract: a rebuild equals a normal open. Trivially true in P1 (every open is already a full WAL replay), and load-bearing the moment an engine-state cache lands. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…exit checklist green Rewrites examples/clog/README.md for the host-integration reader (what clog is, the observe/situation loop, P1 API table, quickstart, config knobs, documented caveats, status/roadmap), adds a runnable examples/clog/examples/quickstart.rs mirrored in the README, and promotes Clog's top-level doctest from no_run to an executing example. Formats the clog crate for the first time (cargo fmt -p clog, no other workspace crate touched) and re-verifies clippy/tests pass after reformatting. Exit checklist green: cargo test -p clog; cargo test -p clog --features test-crash (R1 included); cargo doc -p clog --no-deps (warning-free); cargo clippy -p clog --all-targets [--features test-crash] -- -D warnings; cargo fmt -p clog --check.
Critical:
- wal: a failed append could strand a partial frame mid-log; the next
successful append landed after it and reopen's CRC scan then quarantined
and truncated everything from the tear onward, silently discarding acked
batches. `append` now records the pre-write file length and truncates back
to it on ANY error; if the truncation itself fails the Wal is poisoned and
refuses further appends rather than compounding the damage.
- validate: `scoring_clamp` overflowed `now + 24h` when `now` was within a
day of u64::MAX (reachable - the manual clock saturates there), panicking
the writer under overflow checks. Now saturating.
- actor: an entity's registry display name reached the rendered document
verbatim, so a newline in `EntityRef.name` fabricated document lines. It
now goes through `render::headline` like every other host string.
Important:
- actor/types: INV-5's duplicate skip used `Claim: PartialEq`, whose
`EntityRef` equality ignores `name`, so a re-observe changing only a
display name was swallowed (no commit, registry never updated). New
`claims_identical` compares the name lists alongside, leaving EntityRef's
identity semantics (load-bearing for every index) untouched.
- actor: merge claims stamped occurred_at/observed_at from a clock that
reads 0 on a fresh Manual instance, failing spec 10 validation. Now stamps
`now.max(1)`.
- types/kinds: `Matcher::BodyContains` now carries its own case-sensitivity
flag (spec 5.6) instead of hard-coding case-insensitivity.
- types: Focus boost docs said "additive"; they are multiplicative (5.4).
View::Live and Config.rebuild_on_open docs corrected.
- validate: reject `subject_key: Some("")`/whitespace-only, and
`decay_buckets_per_half_life == 0` (NaN scores) at open.
- validate: focus *value* errors are now Corrupt { "config: ..." } rather
than InvalidFilter, matching the bad-regex ruling; UnknownKind unchanged.
- actor: `mask_header` string-replace replaced by a structural body/scope
comparison, removing the adversarial-scope-name fragility.
Tests: WAL rollback + poison, u64::MAX clamp, entity-name injection,
display-name re-observe, fresh-clock merge, case-sensitive BodyContains,
belief tiers 2/3 in isolation, a real budget-cap + drop-order render test,
min_score on Urgent, and config rejection at open. Goldens unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Boots the agency fixture frozen mid-story (belief competition live, merge pending), re-prints the situation after every command; crash command demos WAL recovery. Store persists at ./clog-repl-data, idempotent reseeding via INV-5. Co-Authored-By: Claude Fable 5 <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
1 participant
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.
What this is
The first milestone pair (M0+M1) of clog — an orientation engine for agentic systems, per
docs/clog-spec-v1.mdas amended bydocs/superpowers/specs/2026-08-15-clog-build-design.md. Hosts write structured claims; clog maintains materialized views incrementally and renders a token-budgeted situation document per scope. Reads are snapshot clones — zero retrieval work at turn time. Clog never calls a model and carries no host semantics.Integration shape (the point of this milestone): install the crate, wire connectors to
observe(), have the agent pullsituation(scope)each turn. Seeexamples/clog/README.mdandcargo run -p clog --example quickstart.Contents
examples/cloglibrary crate (~8k lines): public API (open/observe/retract/revoke_observer/situation/select/merge_entities/advance), pure semantic core (validation, scoring, belief resolution, alias map, rules-tier classifier, deterministic renderer + budgeter), naive oracle engine overimblsnapshot maps, single-writer actor withArcSwappublication, postcard+CRC32 WAL with torn-tail quarantine and write-ahead ordering.imbl, WAL-records-effects,Batch.as_of, masked change detection, alias-map-as-view, canonicalized boosts), resolved fold F-checklist (docs/fold-answers.md), P1 implementation plan.Test evidence
cargo test -p clog: 61 unit + 27 integration + 1 golden + 6 property + recovery + 4 doctests, all green;--features test-crashadds the R1 subprocess crash-point suite (abort after WAL fsync at every batch, reopen equals fresh-fed prefix byte-for-byte).clippy --all-targets -D warningsandcargo fmt --checkclean. Spec invariants are enforced by named tests (INV-1/3/4/5/6/8/9/10/11 in P1 scope); the G1 agency goldens freeze the §5.8 rendering contract (independently hand-recomputed scores).Review process
Built task-by-task with a fresh implementer + independent reviewer per task (5 in-loop fix rounds), then an adversarial whole-branch review: 6 finder lenses, each Critical/Important finding judged by 2 independent refuters. 10 findings survived (3 Critical: WAL partial-frame stranding,
scoring_clampoverflow panic, entity-name render injection) and were fixed + re-reviewed in a single verified wave; 10 were refuted; 15 minors triaged (fixed or recorded).Deferred (by design, P2+)
set_focusat runtime, ticks/decay-bucket recomputation, watches/wakes,recall/kNN + eval harness (garden3d corpus), fold engine adapter, WAL snapshot files + format versioning, single-writer directory locking. Known caveats are documented in the crate README.🤖 Generated with Claude Code