[devday] Reading order: deterministic, proof-derived "read these first" (co-change wired)#331
Open
rpatricksmith wants to merge 16 commits into
Open
[devday] Reading order: deterministic, proof-derived "read these first" (co-change wired)#331rpatricksmith wants to merge 16 commits into
rpatricksmith wants to merge 16 commits into
Conversation
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Wire the written-but-unused namedImport query so extractImports populates names[] for TS/TSX (fixing the names:[] // Simplified stub). The query was also malformed (clause-before-source ordering + name: field required) — fixed in queries.ts for both typescript and tsx. Add analyzers/graph/buildGraph.ts: a deterministic file->file import digraph. Resolves relative + tsconfig-path specifiers (with NodeNext .js->.ts rewrite) to in-repo files; unresolved/external specifiers produce NO edge. Node identity is repo-relative POSIX (matching the symbol index); SymbolEntry is untouched. Persistence to .ana/state/code-graph.json is opt-in via scanProject's new persistGraphTo option, so ana scan stays read-only (no byte-parity regression). init passes its staging state dir so the graph travels through the atomic swap alongside scan.json / symbol-index.json. Deep-tier only, behind the 750 cap. Tests: named-import wiring (incl. no cross-attribution, default/namespace -> no names, Python unchanged); graph resolution (relative/index/alias/external/escape, dedup, self-edge, determinism, fail-soft, .js->.ts rewrite, absolute->relative); a 750-file deep-tier latency gate. Full suite green (3740 pass). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Fuse import-graph PageRank centrality + Slice-1 proof-chain bug-magnet rate +
co-change into one token-budgeted "read these first" list, personalized toward
the active scope when present.
- analyzers/graph/pagerank.ts: ~40-line deterministic power-iteration PageRank
(fixed 40 iterations, damping 0.85, uniform dangling-mass redistribution,
node order from the already-sorted graph) — two runs byte-identical.
- analyzers/reading-order/index.ts: buildReadingOrder() fuses the three signals
into ranked {file, score, reasons[]}, binary-searches the list to a ~1k-token
budget, and boosts in-scope files (1.5x). Every reason states its measured
basis (work items, rework cycles, co-change partner, import centrality) —
never fabricated. Returns null below a 3-edge threshold. resolveImportRelation-
ships() cross-refs co-change rows vs the graph to set hasImportRelationship
(true with edge, false when both nodes in-graph without one, null — never
false — when a file is missing / low-confidence).
- analyzers/reading-order/scope.ts: findActiveScope() reads the single active
.ana/plans/active/<slug>/scope.md "Files affected" line (declines when zero
or many are active — no guessing), rebasing repo-root paths onto node identity.
- scan-engine.ts: build the Slice-2 graph in memory at deep tier (persist still
opt-in), then populate readingOrder + resolve hasImportRelationship. Surface
tier and sparse graphs stay null (byte-parity preserved).
- scan.ts: render the "Start here" card; scaffold-generators.ts: inject the
~15-line Start Here block into project-context.md.
Additive/nullable — existing scan output keeps byte-parity (populates a
currently-null field). Real ana scan surfaces work.ts (68 work items, 14 rework
cycles) and proofSummary.ts (36) fused with centrality; two real-repo scans are
byte-identical. 46 new tests; full suite 3804 pass / 2 skip, lint + test
typecheck clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…lobal PageRank + in-degree floor + CJS) — fixes readingOrder ranking Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…+ orchestrator/entrypoint boost) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…nary → text) The intent-couple pair key used a literal NUL byte (0x00) as its separator, which made git treat the whole file as binary — no diff, no merge, no review. Replace the raw NUL with the \0 escape sequence: byte-identical runtime behavior (NUL is still the separator, the one byte a POSIX path can't contain), but the source is now plain ASCII and reviewable. Also refresh two stale '2 of 202' comments to be count-agnostic. (This commit shows as a binary diff for the last time — the new blob is text; all future diffs are normal.) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…signal The scaffold claimed a three-signal fusion (centrality + rework + co-change) while co-change was computed (intentCouples) but never threaded in — two signals fused, three asserted. Wire intentCouples first-class into the reading-order fusion: honest 'changed together in N verified items' provenance (from coTouchCount/slugs, never a synthetic percentage), gated to >= 2 verified items so one-off co-touches and mega-refactor artifacts don't count, with hidden-coupling detection (co-changed but no shared import edge). The scaffold's co-change clause is now conditional on the signal actually contributing. Also harden two adversarial-review findings in the same path: - Coverage caveat now fires when the import graph was built from the 750-file SAMPLE, closing a honesty hole where mid-size repos (~750-2500 files) showed a partial ranking as whole-repo; polyglot caveat reframed by language (no misleading percentage), percentage clamped to 100. - Co-change top-partner pick now breaks ties on partner path, so the fusion is self-determinizing rather than relying on the caller pre-sorting couples. gitIntelligence.coChangeCoupling stays null on purpose (reserved for a future git-churn path; populating it would demand a fabricated percentage). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…cite The rescan-on-complete refreshes scan.json only — it must NOT persist the import graph, which would write an untracked .ana/state/code-graph.json into the working tree after completion (state/ is only gitignored once ana init has run). Nothing reads code-graph.json on the live path, so the marginal freshness isn't worth churning the tree; it refreshes on the next ana init. Also corrects the stale 'scan.ts:470' worktree-guard cite in the hook comment. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Fixturize proof-history count assertions: the 'verified premises' test pinned exact touch counts (68/42/36) against the LIVE, ever-growing proof_chain.json, so any merge that completed work broke it (work.ts 68 -> 69 on this merge). Lock exact counts against a FROZEN synthetic fixture instead; the live-data tests now assert only drift-robust invariants (rank, gate, determinism). - Co-change: integration tests drive it end-to-end through scanProject (the live path always passed []) — proving it fires, gates one-offs, flags hidden coupling, stays null-safe on a degenerate chain, and a non-tautological order-flip test (co-change lifts d.ts above a.ts against the path tie-break, failing if the score term is removed — proving it changes ORDER, not reasons). - New coverage: sampled-graph caveat, polyglot wording, partner-pick determinism under shuffled input, conditional co-change subtitle, and a fusion timing gate. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The reading-order differentiator (deterministic, proof-derived 'read these first') was invisible in docs and the CHANGELOG was untouched by the whole feature. Add a 'Where to start reading' section to the scan concept doc (three signals, hidden coupling, determinism, the honest coverage caveat — no numbers) and a comprehensive [Unreleased] CHANGELOG entry covering the reading order, proof-history risk map, import graph, rescan-on-complete, and the co-change honesty fix. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
CI (clean checkout) caught a type error a stale local incremental tsc cache
hid: the perf-gate edge literals were {from,to} but ImportEdge requires names.
Annotate as ImportEdge[] and add names: [] — matches the edge() helper used by
the other reading-order tests.
Co-Authored-By: Claude Opus 4.8 (1M context) <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 ships
A deep-scan reading order — a deterministic, proof-derived "read these first" ranking written into the
ana scancard and the project-context scaffold. It fuses three measured signals:Plus the proof-history risk map, the TS/JS import graph + PageRank, and a non-blocking rescan-on-complete hook.
The headline fix: co-change is now real, not asserted
The branch previously computed co-change (
intentCouples) but never threaded it into the fusion, while the scaffold claimed "Fused from import centrality, proven rework risk, and co-change." Two signals fused, three asserted — a by-construction-honesty violation in a product whose whole thesis is honesty.Co-change is now wired first-class:
Evidence it helps (real before/after on this repo):
work.tsrises #6→#1 (it's the genuine top hotspot), the pipeline core cluster (work/proof/artifact) surfaces correctly with hidden-coupling flags, and both promotions are legitimate high-churn core files — no test/fixture noise promoted.Also in this PR (quality + honesty hardening)
0x00separator that made the file git-binary (unreviewable); escaped to\0, byte-identical runtime, now plain text.proof_chain.json(broke on this merge: 68→69). Exact counts now lock against a frozen synthetic fixture; live-data tests assert only drift-robust invariants.Validation
main(post-config [devday-config] Optional configurability — honest Tier-1 core (Tier-2 excised) #330); clean,work.tscomposition verified line-by-line (main's verdict gate + this branch's rescan hook are non-interacting).🤖 Generated with Claude Code