Add shrink-only context budget gate - #622
Conversation
|
The latest updates on your projects. Learn more about Vercel for GitHub. 1 Skipped Deployment
|
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d55a789acd
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review: PR #622 — Add shrink-only context budget gate
Recommendation: REQUEST CHANGES
Summary
This PR adds a deterministic CI gate (tools/check-context-budget.mjs) enforcing a shrink-only byte budget over the always-loaded CLAUDE.md + .claude/rules/core.md, guards the root-import and unconditional-rules allowlists, wires a standalone Context Budget job into guards.yml, and adds nine hermetic decision-path tests. The tool's --check logic itself is sound and well-covered. The defect is in the baseline the PR commits alongside it: tools/context-budget.json does not match the actual byte size of the files it certifies — including .claude/rules/core.md, a file this PR never touches.
Findings
High
Seeded tools/context-budget.json does not match the actual byte counts of the files it certifies
- Location:
tools/context-budget.json:1(baseline),CLAUDE.md,.claude/rules/core.md - The committed baseline records
CLAUDE.md=18950,.claude/rules/core.md=3443(total 22393). Verified directly viagit cat-file -sagainst the PR's merged blobs:CLAUDE.mdis actually 18864 bytes andcore.mdis actually 3380 bytes (total 22244) — 149 bytes lower than the seeded baseline. core.mdis untouched by this PR (git diffbetween the pre-PR base and the PR commit shows zero changes to that file), yet its baseline entry is inflated by exactly 63 bytes — which exactly matches the file's own line count (63 lines viawc -l), the signature of one stray\rper line. Since the file's content is unchanged, the only explanation is that the baseline was seeded from a CRLF-materialized local checkout..gitattributeshas notext eol=lfrule forCLAUDE.mdor.claude/rules/*.md(only.claude/workorders/*.mdand a few others are covered), so a--write-baselinerun on a CRLF checkout silently inflates every measured byte count — the exact class of bug this repo has hit before (per the.gitattributescomment about CRLF materializing on a fresh Windows clone).- Risk: CI checks out on Linux with no CRLF conversion, so the real enforced measurement (22244) is below the inflated baseline (22393) — the check passes today, but only by accident, with ~149 bytes of unearned, undocumented slack baked into the exact commit whose entire purpose is byte-for-byte shrink-only precision. Any future growth up to that slack goes undetected, so the shrink-only guarantee is false from the seed forward.
- Fix: Add
CLAUDE.md text eol=lfand.claude/rules/*.md text eol=lfto.gitattributes(matching the existing.claude/workorders/*.mdprecedent), then re-runnode tools/check-context-budget.mjs --write-baselinefrom an LF-normalized checkout and recommittools/context-budget.jsonwith the corrected numbers (CLAUDE.md 18864, core.md 3380, total 22244) before merge.
What's good
check-context-budget.mjs's--checklogic (allowlist checks, delta computation, JSON/human output, malformed-baseline handling) is clean and thoroughly exercised by nine well-targeted hermetic decision-path tests.- The
context:reseedescape hatch mirrors the repo's existingratchet:reseed/parity:exemptpattern. - Docs kept in lockstep:
tools/README.mdand theCLAUDE.mddocs registry both updated in the same commit.
Deferred — not applicable to this diff
No apps/web/**, apps/mobile/**, packages/shared/src/types/*, endpoints.ts, or orbit-api files changed, so design/parity/i18n/contract/security/backend-hard-rules dimensions don't apply. Lint/type/test are covered by this PR's own required CI checks (Build / Unit Tests / SonarCloud), per this workflow's Phase-6 skip.
| Severity | Count |
|---|---|
| Critical | 0 |
| High | 1 |
| Medium | 0 |
| Low / Info | 0 |
There was a problem hiding this comment.
Code Review: PR #622 — Add shrink-only context budget gate
Scope: PR #622 in thomasluizon/orbit-ui-mobile (ORB-117)
Recommendation: NEEDS WORK
Summary
This PR ships tools/check-context-budget.mjs, a new CI gate that enforces a shrink-only
byte budget on the always-loaded context surface (root CLAUDE.md + unconditional
.claude/rules/*.md), plus structural allowlists for sibling @ imports and unconditional
rule files, wiring into guards.yml, tools/test-tools.mjs, tools/README.md, and root
CLAUDE.md's own docs. The tool's logic, git-based target-branch comparison (fixed in the
second commit to compare against the actual PR base SHA instead of a mutable branch/local
file), and the eleven hermetic decision-path tests are sound and well covered. However, the
seeded baseline this PR commits, the entire deliverable of "seed the baseline", does not
match the actual byte count of the files it describes, verified directly against the
committed git blobs. This is a Correctness/Harness-evidence defect that blocks merge; see
High below. No parity, i18n, contract, design, or backend-hard-rule surfaces are touched
(tooling/CI-only diff), so those dimensions are N/A.
Findings
Critical
None.
High
[HIGH] Committed context-budget baseline does not match the files it claims to measure
· dimension: Correctness (#1), reinforced by Harness changes need EXECUTED evidence (#15)
· location: orbit-ui-mobile/tools/context-budget.json:1-7 (and the "Seeded baseline" block in the PR description)
· issue: The committed baseline records CLAUDE.md: 18950 bytes and .claude/rules/core.md: 3443 bytes
(total 22393), but the actual blobs committed in this same PR are 18864 and 3380 bytes
(total 22244), a 149-byte overstatement. Verified independently two ways against the raw git
objects (bypassing any working-tree/checkout filter):
git rev-parse 71d8bc3:CLAUDE.md -> 3feb01627e13dfe6e6e42e69611e543638a58cf2
git cat-file -s 3feb01627e13dfe6e6e42e69611e543638a58cf2 -> 18864
git rev-parse 71d8bc3:.claude/rules/core.md -> 009fac02b053fa0457b24ef214558f8a297746df
git cat-file -s 009fac02b053fa0457b24ef214558f8a297746df -> 3380
git diff d55a789a 71d8bc3a -- CLAUDE.md .claude/rules/core.md tools/context-budget.json is empty,
so the mismatch isn't later-commit drift within the PR, the baseline was stale from the very
commit (d55a789a) that both edited CLAUDE.md and seeded the baseline in the same commit.
An independent skeptic pass (fresh subagent, told to try to refute this) confirmed the same
numbers via the same blob-hash method and could not find an alternate-path or
BOM/CRLF/tool-semantics explanation, check-context-budget.mjs's byteCount() is a plain
readFileSync(path).byteLength on CLAUDE.md and .claude/rules/core.md only, with no
normalization that could account for 149 bytes.
· risk: The gate is shrink-only (fails only when measured bytes exceed the baseline). Because the
seeded baseline is 149 bytes higher than the true byte count, every future PR gets up to 149
bytes of real, unflagged growth in root CLAUDE.md + core.md before the Context Budget job
ever fires, precisely the "changes... can no longer increase the committed byte budget
silently" guarantee the PR's own "Developer impact" section claims to deliver. The pasted
"Seeded baseline" JSON and the "Review round 1 validation" --check --json transcript in the
PR body are themselves not accurate evidence of the final shipped state (execution happened,
but not against the content that actually merged), which is exactly what rubric dimension 15
exists to catch: a claim that a tool works, without a command that ran against the real
state, is a review finding.
· fix: Regenerate the baseline against the final committed content and commit the corrected
file: node tools/check-context-budget.mjs --write-baseline, which should produce
{"bytes": 22244, "files": {"CLAUDE.md": 18864, ".claude/rules/core.md": 3380}}. Update the
"Seeded baseline" block in the PR description to match, and re-paste a --check --json
transcript run against the final commit (not an intermediate draft).
· reference: CLAUDE.md rule 1 (root cause); rubric dimension 1 (Correctness) and dimension 15
(Harness changes need EXECUTED evidence).
Medium
None concretely actionable beyond the High above.
Low / Info
None posted (signal gate: Low/Info are not posted on a PR review).
Subagents
| Agent | Verdict |
|---|---|
| parity-checker | N/A, no apps/web/** or apps/mobile/** file changed |
| i18n-syncer | N/A, no user-facing strings or i18n JSON changed |
| contract-aligner | N/A, no packages/shared/src/types/* / endpoints.ts / orbit-api DTO changed, and orbit-api is untouched |
| security-reviewer | N/A, no orbit-api code changed |
| design-reviewer | N/A, no apps/* or landing-page UI file changed |
Validation
| Check | Result |
|---|---|
| Lint | PASS (CI: Lint job, PR Tests, SUCCESS) |
| Type check | PASS (CI: Type Check job, PR Tests, SUCCESS) |
| Tests | PASS (CI: Unit Tests job, PR Tests, SUCCESS) |
| Build (api) | N/A, orbit-api untouched |
tools/test-tools.mjs + .claude/hooks/test-hooks.mjs |
PASS (CI: Harness Execution job, Guards, SUCCESS) |
check-context-budget.mjs --check (this PR's own gate, against itself) |
PASS (CI: Context Budget job, Guards, SUCCESS) — expected given the baseline is higher than the true value (see High finding); a passing gate here does not contradict the finding |
node execution was unavailable directly in this review sandbox, so validation above is
sourced from the PR's own live CI check-run results (gh pr view 622 --json statusCheckRollup)
rather than a local re-run; the High finding itself was verified independently via raw
git cat-file / git rev-parse (no node required), cross-checked by an adversarial skeptic
subagent.
Deferred — N/A dimensions & files not verdicted
- Dimension 8 (DESIGN.md / AI-slop): N/A, no
apps/*UI files touched. - Dimension 9 (Parity): N/A, no
apps/web/**orapps/mobile/**files touched. - Dimension 10 (i18n): N/A, no locale JSON or user-facing strings touched.
- Dimension 11 (Contract drift + backward-compat): N/A, no
packages/shared/src/types/*,
endpoints.ts, or orbit-api DTO touched; no backward-compat candidates in this diff. - Dimension 12 (Security): reviewed inline (no injection surface,
spawnSyncuses arg
arrays, not shell strings;CONTEXT_BUDGET_BASE_REFcannot reach a shell); no findings. - Dimension 13 (Backend hard rules): N/A, orbit-api not touched.
- Dimension 14 (FEATURES.md parity): N/A, no user-facing feature surface changed (internal
CI tooling only). - Every changed file (
guards.yml,CLAUDE.md,tools/README.md,
tools/check-context-budget.mjs,tools/context-budget.json,tools/test-tools.mjs) was
read in full and given a verdict; nothing left unexamined.
What's good
- The git-based target-branch comparison (comparing against the PR's actual base SHA via
git show <ref>:<path>/git ls-tree, not a local file or a mutable branch name) correctly
defeats a local--write-baselineattempt to hide real growth from CI, confirmed by the
dedicated test "a regenerated working baseline cannot hide growth from the target branch". - Fail-closed behavior throughout: an unresolvable target ref, a malformed baseline, and an
unallowlisted import/rule file all exit non-zero rather than silently passing. - Strong, concrete hermetic test coverage: eleven distinct decision paths in
tools/test-tools.mjs, matching the Harness Execution contract exactly (dimension 15
satisfied on the coverage axis, independent of the baseline-data defect above). - Zero comments in the new script; all self-documenting via naming, clean under the strict
comment policy. spawnSyncis used with argument arrays throughout (resolveTargetRef,readComparisonBaseline),
not shell string interpolation, soCONTEXT_BUDGET_BASE_REFcannot become a command-injection
vector even though it flows from a GitHub Actions expression into an env var.
Recommendation
Regenerate and commit an accurate tools/context-budget.json against the PR's final content
(node tools/check-context-budget.mjs --write-baseline), correct the PR description's "Seeded
baseline" and validation transcript to match, then this is ready to merge, the tool's logic,
CI wiring, and test coverage are otherwise solid.
|
@codex review |
|
Codex Review: Didn't find any major issues. Swish! Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
There was a problem hiding this comment.
Code Review: PR #622 — Add shrink-only context budget gate
Scope: PR #622 in thomasluizon/orbit-ui-mobile (ORB-117)
Recommendation: APPROVE
Summary
This is a pure tooling/CI-infra PR: a new tools/check-context-budget.mjs gate that enforces a shrink-only byte budget for the always-loaded CLAUDE.md + unconditional .claude/rules/*.md files, plus a seeded baseline, a new Context Budget guards.yml job, an LF .gitattributes pin for the two enforced files, harness coverage for all decision paths, and doc updates (CLAUDE.md, tools/README.md). No apps/* or orbit-api files are touched, so the platform-parity, i18n, contract-drift, backend-security, DESIGN.md, and FEATURES.md dimensions are all N/A.
Two prior claude reviews on this PR (CHANGES_REQUESTED) both flagged the same High finding: the committed tools/context-budget.json baseline (18950/3443/22393) did not match the true byte count of the CRLF-free committed blobs (18864/3380/22244), a 149-byte overstatement traced to a CRLF-materialized local checkout used to seed the baseline. This review independently re-verified the current state: git show HEAD:tools/context-budget.json now records {"bytes": 22244, "CLAUDE.md": 18864, ".claude/rules/core.md": 3380}, and git show HEAD:CLAUDE.md | wc -c → 18864, git show HEAD:.claude/rules/core.md | wc -c → 3380 — an exact match. The .gitattributes LF pin for CLAUDE.md and .claude/rules/*.md is present in the diff, and the git log shows the two follow-up commits ("Fix context budget LF baseline", "Fix context budget base comparison") that resolved this. The High finding from both prior rounds is resolved; no new issues found.
Findings
Critical
None.
High
None. (Prior High — stale/CRLF-inflated baseline — is resolved; verified independently against the current committed blobs, see Summary.)
Medium
None.
Low / Info
tools/check-context-budget.mjs— the pass/fail verdict is deliberately total-only: a file that individually grows can still pass if another enforced file shrinks by an equal or greater amount. This is an explicit, PR-documented decision, not an oversight — flagging for visibility only.- The
context:reseedlabel (guards.ymlcontext-budgetjob) skips the CI check but doesn't itself force a--write-baselinecommit; a reseeded PR that forgets to commit the new baseline leavesmain's baseline stale. This mirrors the existingratchet:reseedpattern in the same file, so it's consistent with established convention rather than a new gap.
Subagents
| Agent | Verdict |
|---|---|
| parity-checker | N/A — no apps/web/** or apps/mobile/** file changed |
| i18n-syncer | N/A — no user-facing strings or shared i18n JSON changed |
| contract-aligner | N/A — no packages/shared/src/types/* / endpoints.ts, no both-repo change |
| security-reviewer | N/A — no orbit-api code changed |
| design-reviewer | N/A — no apps/* or landing-page UI file changed |
Validation
| Check | Result |
|---|---|
| Lint / Type check / Tests | PASS (this PR's own required CI checks — Build / Unit Tests / SonarCloud — per this workflow's Phase-6 skip) |
Baseline accuracy (tools/context-budget.json vs actual blobs) |
PASS — independently re-verified in this session, exact match (18864 / 3380 / 22244) |
Harness (tools/test-tools.mjs) |
PASS per PR body / prior review verification; test wiring for check-context-budget.mjs confirmed present and registered |
Hooks (.claude/hooks/test-hooks.mjs) |
PASS per PR body |
Deferred — N/A dimensions
- Parity, i18n, contract drift/backward-compat, backend security/hard rules, DESIGN.md/AI-slop, FEATURES.md parity — all N/A, no
apps/*ororbit-apifiles in the diff.
What's good
- The bootstrap-vs-enforce split is correctly asymmetric: first-run reads the working-tree baseline, every subsequent check reads the baseline from the git target ref, so a contributor cannot defeat shrink-only enforcement via a local regeneration.
- The
.gitattributesLF pin for exactly the two enforced files is the right fix for the CRLF byte-count instability the prior review rounds caught, and it's independently confirmed correct against the committed blobs. - Structural findings (new root imports, new unconditional rules) are evaluated independently of shrink/grow and block even a
--write-baselinerun. - Harness coverage is thorough across the tool's decision paths.
Recommendation
Nothing blocks merge. Both prior High findings are resolved and independently reverified. The two Info-level notes above don't need action in this PR.
There was a problem hiding this comment.
Code Review: PR #622 — Add shrink-only context budget gate
Recommendation: APPROVE
Summary
Adds tools/check-context-budget.mjs, a deterministic, dependency-free CI gate enforcing a shrink-only byte budget for the always-loaded context (CLAUDE.md + unconditional .claude/rules/*.md), plus a structural allowlist for root @ imports and unconditional rule files. Wires it into guards.yml behind a context:reseed label escape hatch, pins CLAUDE.md and .claude/rules/*.md to LF via .gitattributes so a Windows checkout measures the same bytes CI does, seeds tools/context-budget.json, and adds thorough test-tools.mjs coverage. Traced line-by-line against tools/CONVENTIONS.md and the actual current .claude/rules/ state (which file is unconditional vs. paths:-scoped); no correctness defect found.
Findings
Critical / High / Medium
None.
Low / Info
--check doesn't itself flag a diff that removes tools/context-budget.json while the target branch still has one (tools/check-context-budget.mjs, readComparisonBaseline() / the --check path in main()). --check compares current bytes against the target branch's committed baseline; it never asserts the current branch still carries the file when the target has one. A PR that deletes tools/context-budget.json (bad merge/rebase, stray cleanup) would still pass as long as the byte delta is ≤ 0, silently merging the deletion. This is self-limiting though: the next PR against that main hits the bootstrap path, finds the file missing from both target and its own working tree, and readWorkingBaseline() correctly fails closed with exit 2. So the real blast radius is one merged PR, not a permanently disabled gate. Worth having --check name the deletion directly rather than relying on the next unrelated PR's failure to surface it — not blocking.
Subagents
All five (parity-checker, i18n-syncer, contract-aligner, security-reviewer, design-reviewer) are N/A: the diff touches no apps/* UI, no packages/shared/src/types/*/endpoints.ts, no orbit-api code, and no i18n strings — this is a pure CI-tooling / internal-DX change to orbit-ui-mobile only.
Validation
This review's sandbox could not obtain approval to re-execute npm run lint/type-check/test or node tools/test-tools.mjs directly. In place of trusting the PR body's transcript outright, the byte totals underpinning the whole PR were independently re-derived from the merge commit:
$ git show HEAD:CLAUDE.md | wc -c
18864
$ wc -c .claude/rules/core.md
3380
Both match the PR body's raw-blob-verification section and tools/context-budget.json's 22244 total exactly. The PR body's own transcript is a full command+output log (not a bare claim), which is what D9/rubric dimension 15 requires of a harness change.
What's good
- The structural allowlists mean
--write-baselineitself refuses to run if the import/rule allowlist has silently grown (structuralFindingsis checked before either the write or delta path) — matchestools/CONVENTIONS.md's "verdict computed from artifacts on disk" gate-tools rule. resolveTargetRefcorrectly distinguishes "target branch unreachable" (fail closed, exit 2) from "target branch reachable but lacks the baseline file" (first-run bootstrap) viagit ls-tree.- The
.gitattributesLF pin is a genuine root-cause fix (Code Standard 1) for a real cross-OS byte-counting hazard, not a workaround inside the checker. - Test coverage is unusually complete: over/under budget, no-rewrite-on-pass, baseline-regeneration-can't-hide-growth, first-run bootstrap, unfetched-target fail-closed, unallowlisted-import-even-when-target-absent, new-unconditional-rule, scoped-rule-stays-out, missing/present sibling imports (both asserted), malformed baseline, and a
--helpcompleteness assertion.
Recommendation
Approve as-is. The one Low/Info note is a future polish, not a blocker.
|
@codex review |
|
|
Codex Review: Didn't find any major issues. Nice work! Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
There was a problem hiding this comment.
Code Review: PR #622 — Add shrink-only context budget gate
Scope: PR #622 in thomasluizon/orbit-ui-mobile (ORB-117)
Recommendation: APPROVE
Summary
Adds tools/check-context-budget.mjs: a dependency-free CLI enforcing a shrink-only total byte budget for always-loaded context (CLAUDE.md + unconditional .claude/rules/*.md), rejecting unlisted growth of the root @ import allowlist and the unconditional-rules allowlist, and reporting (not enforcing) resolvable sibling CLAUDE.md files. Wires it into a new Context Budget job in .github/workflows/guards.yml, seeds tools/context-budget.json, pins CLAUDE.md/.claude/rules/*.md to LF in .gitattributes (closing a real Windows-vs-CI byte-count hazard the repo already hit once), and documents the tool in CLAUDE.md's docs registry and tools/README.md.
Changed files: .gitattributes, .github/workflows/guards.yml, CLAUDE.md, tools/README.md, tools/check-context-budget.mjs, tools/context-budget.json, tools/test-tools.mjs.
I traced the tool's full control flow (arg parsing, measurement, baseline parse/compare, target-ref resolution, bootstrap vs steady-state, write-baseline) against every fixture in the new contextBudgetCases in tools/test-tools.mjs and found it internally consistent — no correctness defect.
Prior review history on this PR
This PR went through two rounds of CHANGES_REQUESTED (addressed by the "Fix context budget base comparison" and "Fix context budget LF baseline" commits), followed by two prior APPROVED reviews after those fixes landed. No PR-content commits have landed since the last approval — the subsequent merge commits (9a10e8d6, 480946a2) only carried forward unrelated changes already on main (#624); git diff between the last-reviewed commit and the current PR head shows zero changes to files this PR actually touches. This review confirms that assessment independently rather than re-litigating resolved findings.
Findings
Critical: None. High: None. Medium: None.
Low / Info:
[INFO] Context Budget CI job self-compares on push to main
- dimension: SOLID/clean-arch (informational)
- location:
.github/workflows/guards.yml(context-budgetjob) - issue: The job's
ifexcludespull_requestfrom thecontext:reseed-labeled case but still runs onpushtomain. There,CONTEXT_BUDGET_BASE_REFresolves toorigin/main, which a push-triggeredactions/checkout --fetch-depth 0updates to the very commit just checked out — so the job compares that commit's baseline against itself, a guaranteed pass. Sincemainis protected (no direct push), every commit reachingmainalready passed this same check at PR time, so this is a wasted (never-failing) CI job onpush, not a masked regression. - fix: Optional — gate to
pull_requestonly, mirroring theratchetjob'sif: github.event_name == 'pull_request' && ...pattern.
Subagents
All five parity/contract/security/i18n/design subagents gated N/A: no apps/web/**/apps/mobile/** file, no user-facing string, no packages/shared/src/types/*/DTO, no orbit-api code, no UI file changed. Cross-repo dimensions (contract-aligner, orbit-api side of backward-compat guard) not verifiable in this CI job — orbit-api is not checked out here, and this PR's diff doesn't touch any cross-repo surface regardless.
Validation
Per this workflow's CI adaptations, Phase 6 (/validate) was skipped — Build, Unit Tests, and SonarCloud already run as separate required checks on this PR and are green. All CI checks are green, including the new Context Budget job itself (self-validating).
What's good
- Structural allowlists (imports, unconditional rules) are closed sets that
--write-baselinecannot launder growth into —main()fails closed on structural findings before it ever reaches the write path. readComparisonBaselinecorrectly distinguishes "target branch has no baseline yet" (bootstrap) from "target branch is simply unfetched" (fail closed, exit 2, actionable message).- Test coverage is decision-path-shaped (bootstrap, regenerated-baseline-can't-hide-growth, unfetched-target fail-closed, malformed-baseline, scoped-rule exemption), not happy-path-only.
- The LF-pinning fix in
.gitattributescloses a real cross-platform byte-count hazard this repo already hit once (ccc1849f).
Recommendation
No Critical/High findings, all CI green, and no PR-content changes since the last approved review. Safe to merge as-is; the one Low/Info CI-redundancy note is optional cleanup, not a blocker.
1 participant
ORB-117
Adds a deterministic, dependency-free gate for the always-loaded context budget. The checker enforces a shrink-only byte total for repository-visible files, rejects additions to the root import and unconditional-rule allowlists, and reports resolvable sibling context without enforcing it.
The gate compares pull request measurements with the target branch baseline.
CLAUDE.mdand.claude/rules/*.mdare pinned to LF so Windows baseline writes measure the same bytes Git and CI store.Developer impact
Changes to root
CLAUDE.mdor unconditional.claude/rules/*.mdfiles can no longer increase the committed byte budget silently. New root imports and unconditional rules also fail structurally even when their target content is unavailable in CI.Seeded baseline
Final output from commit
ccc1849f:{ "files": { "CLAUDE.md": 18864, ".claude/rules/core.md": 3380 }, "enforcedBytes": 22244, "estimatedTokens": 5561, "imports": [ "@../orbit-api/CLAUDE.md", "@../orbit-landing-page/CLAUDE.md" ], "unconditionalRules": [ ".claude/rules/core.md" ], "unexpectedImports": [], "unexpectedRules": [], "siblingFiles": [], "fullSessionBytes": 22244, "fullSessionEstimatedTokens": 5561, "baselineBytes": 22244, "baselineSource": "refs/remotes/origin/main (working tree bootstrap)", "deltaBytes": 0, "fileGrowth": [], "structuralFindings": [], "status": "ok" }Raw blob verification
The recorded file counts equal the raw Git blob sizes, and
18864 + 3380 = 22244.Validation
Decisions taken unattended
origin/main, thenmain; if neither ref resolves, the tool exits 2 and requests target history.git ls-tree. Only the former may use the working tree baseline for first-run bootstrap.CLAUDE.mdand.claude/rules/*.md, then freshly restored the measured files before running--write-baseline.tools/context-budget.jsonthrough the tool rather than editing its numbers by hand.