v0.14.0

Upgrading to v0.14.0

Default changes

• agent-scratch: exempt_paths without the bootstrap → exempt_paths gains .claude/hooks/discipline-bootstrap.sh (looser). The script hook install --agent claude-code writes is shared project configuration, like .claude/settings.json; instruction-smuggling still reports a change to it. Restore: [gates.agent-scratch] exempt_paths = [".claude/settings.json", ".cursor/hooks.json", ".cursor/mcp.json", ".aider.conf.yml"].
• pii: every ~/.<agent> path reported → a tool's own configuration location allowed (agent_config_standard_paths = true) (looser). Documenting where a tool reads its settings or hooks is not a leak of a maintainer's setup; the personal content under the same directory is still reported. Restore: [gates.pii] agent_config_standard_paths = false.

Behaviour changes

• version-lockstep, manifest-sync (stricter): DISCIPLINE_REPLAY_CASE, which lets discipline replay skip a group or rule whose file neither side has, is honoured only when the base is the commit replay builds; set in a CI job or by hand, a missing file stays a configuration error (exit 2). Migration: None, unless a pipeline set DISCIPLINE_REPLAY_CASE: remove it.
• dependency-delta (stricter): uv.lock, composer.lock and Gemfile.lock are read entry by entry, as the other lockfiles are: an entry from a new source, a dropped integrity hash and a deleted lockfile are findings, and the package count note is filled. Before, the gate never collected these files, though their parsers existed. Migration: A Python (uv), PHP or Ruby project can receive lockfile findings it did not before.
• vacuous-tests (looser): New directive allow-vacuous-test: <test> <reason> (also discipline:allow(vacuous-tests)) lifts every finding on that one new test; before, no directive could, and only exempt_paths, a helper declaration or the baseline did. Migration: None.
• replay (changed (replay --json shape)): --json: gates whose overrides fail_on_overrides refused are listed in a case's own refused_overrides field (with the actor the case was checked as) and in the summary's refused_overrides_by_gate, instead of in blocking_gates with the reason in detail. Refusal is now judged from the configuration's override policy, so a change blocked by both an error and a refused override names both. Migration: A consumer reading refused overrides from blocking_gates or detail reads refused_overrides.
• msrv (changed (exit 1 becomes 2)): A command that cannot run (not found, cannot start, over the timeout) and a Cargo.toml that cannot be read exit 2 (could not check) instead of reporting a finding; a command that runs and exits non-zero is still a finding. The note for a declaration alone says MSRV declared, not verified. Migration: A pipeline that treated exit 1 as the MSRV build failing sees exit 2 when the toolchain or command is missing.
• assertion-reduction, vacuous-tests (looser): An extra_assert_macros entry written with its ! (assert_matches!) names the same macro as without it; before, it silently never matched, so the tests asserting through it were read as having no assertion. Migration: None.
• suppression-delta (looser): Java: @SuppressWarnings named in a comment or a Javadoc {@code ...} is no longer a new suppression; only the annotation node counts, as documented. Migration: None.
• configuration, report (additive): Renamed configuration keys are supported: an old name listed in KEY_ALIASES is read as the new one with a note in the report's new deprecations list (deprecated: ... in text, **Deprecated:** in the step summary), and setting both names is a configuration error. No key is renamed yet, so no configuration changes behaviour. Migration: None.
• report, replay (additive): JSON Schemas of the check --format json report and the replay --json summary ship as discipline.report.schema.json and discipline.replay.schema.json, generated by discipline docs and checked by docs --check; a test pins every field and validates real output. The output itself does not change. Migration: None.
• report, agent surfaces (changed (agent-prompt heading)): Every finding has a registered code, gate/code: a new code field on each JSON report violation, the [gate/code] heading in the agent-prompt report (hooks, discipline mcp), and a Findings: list in discipline explain, which also accepts a code or a [gate/code] line. Titles, fingerprints and SARIF rule ids are unchanged. Migration: A script matching [gate] in agent-prompt text matches [gate/.
• baseline, report, SARIF, GitLab (changed (SARIF rule ids, GitLab fingerprints, baseline format)): Fingerprint version 2 keys on the finding's code (v2:gate/code:path:hash), not its title: baseline --write writes version = 2 with rule = "gate/code". A version-1 file still matches, with a deprecated: line, until discipline baseline --migrate rewrites it; config-integrity accepts that migration alone without a directive. Each finding in --format json gains fingerprint; SARIF ruleId is the finding code (one rule per code, gate as a tag) with partialFingerprints; the GitLab Code Quality check_name is the code and its fingerprint is the baseline fingerprint. Source-parse findings are coded under assertion-reduction whichever AST gate reports them. Migration: Code scanning and GitLab re-open their alerts once under the new ids; run discipline baseline --migrate in a change of its own.
• agent surfaces (changed (agent-prompt text)): The Repair: line of the agent-prompt report (hooks, discipline mcp) is chosen per finding code, then per gate: a dropped lockfile hash, a loosened constraint, a sleep or a retry added to a test, a missing benchmark baseline and 40 other kinds get their own repair instead of their gate's. No repair tells the agent to regenerate golden output, and repairs named for gate ids that do not exist are gone. Migration: None.
• miri, sanitizers (stricter): A run that could not start (the tool not found, over its timeout) or that reports its toolchain unavailable exits 2 (could not check) whatever the directives say; before, a start failure was a finding an execution, toolchain or nightly waiver lifted, and a waiver also lifted an unavailable toolchain. allow-miri: / allow-sanitizers: now lift only findings from a run that happened. Migration: A job without the toolchain disables the gate in its configuration instead of waiving it.
• findings, report, agent surfaces (changed (titles)): Finding titles are display text: 80 are reworded to one grammar (Title Case, ASCII, no data, one verb per meaning; docs/GATES.md "Finding Codes"), and the 25 that carried data or repeated their message get a fixed title, the data staying in the message. Titles that named a waiver ("Without Directive", "Without Override") no longer do. The code, fingerprint and SARIF rule of every finding are unchanged, and a fingerprint-version-1 baseline still matches the renamed findings under their old titles. GATES.md lists every code with its title. Migration: A consumer matching a title matches the finding's code.
• check (changed): A run that could not check (exit 2) prints the JSON report on stdout under --format json (before, stdout was empty) and writes the same report to --json-out and a JSON --output-file: no outcomes, and a could_not_check object with the reason (configuration, baseline, repository, tool-missing, tool-timeout, toolchain-unavailable, forge, gate, internal), the gate that could not run and the error. Before, --json-out held a single engine outcome with the error as its violation; JUnit, SARIF and GitLab reports still do. Every JSON report and replay summary starts with schema_version (1). The error on stderr is unchanged. Migration: A consumer that read the error from outcomes[0] of an exit-2 --json-out report reads could_not_check.detail. One that treated any stdout under --format json as a completed check reads the exit code first.
• replay (changed): could_not_check_by_reason is keyed by the reason code the child check reported (forge for a change whose merged pull request could not be read), no longer by the last line of its error; each case that could not be checked carries reason, and the text summary lists each change with its error under its reason. Migration: A consumer that matched error text in the keys reads cases_detail[].detail.
• MCP check_diff (additive): A check that could not run carries reason and gate in structuredContent, beside status: could_not_check. Migration: None.
• MCP check_diff (additive): structuredContent carries schema_version and, when the check ran, findings: each finding's code, severity, title, location, message, repair and fingerprint, without the remediation. The tool declares it as its outputSchema. Migration: None.
• agent-prompt, hook run, MCP (changed): Each problem is quoted in a fenced block one backtick longer than any backtick run in it (- Problem: is followed by the block instead of the text), and so is the error of a run that could not check; titles and locations are kept to one line. Directive redaction is case- and spacing-insensitive (ALLOW-SWALLOW:, discipline : allow (x)), and the bare words removes and deletes are no longer redacted without their colon. Migration: A consumer that parsed - Problem: <text> reads the fenced block after - Problem:.
• hook run (stricter): A stop let through at a loop guard (stop_hook_active, agy's third block) still runs the check, and when the change has findings or could not be checked says so on stderr; the exit code and stdout are the agent's pass, as before, so the loop still ends. Migration: None.
• baselines (stricter): A version-2 fingerprint hashes the finding's anchor (a test, dependency, key, counter or commit name) with its source line, and never its message: a finding with no line and no anchor hashes its code and path only. Anchored: deletion-rationale/test-removed-without-rationale, vacuous-tests/vacuous-test-added, the dependency-delta findings on a manifest entry, config-integrity/gate-weakened, toolchain-config/toolchain-config-weakened, bench-regression/counter-regressed and counter-regressed-unapproved-arm, issue-link/directive-in-subject-line. Two findings reported at identical lines (two tests' @Test attribute) no longer share a fingerprint, so baselining one no longer hides the other. Migration: Version-2 baselines written before this change are regenerated with discipline baseline --write; version-1 baselines and --migrate are unchanged.
• hook install (stricter): --agent agy writes the Stop handler directly under the named hook, as agy's hook guide requires; the file written before (the handler inside a matcher / hooks group) was never run by agy. --agent copilot adds apply_patch to the postToolUse matcher, the edit tool some models use. Migration: Re-run discipline hook install in a checkout without the old file (it never rewrites one), or edit .agents/hooks.json and .github/hooks/discipline.json to match.
• explain (changed): allow-nul: is listed under assertion-reduction, the gate that codes the NUL-byte finding, instead of vacuous-tests. Migration: None.
• hook install, hook run (additive): New hook install --agent copilot --user writes hooks/discipline.json in the Copilot home directory (.copilot in the user's home, or COPILOT_HOME), which Copilot CLI loads in any folder, trusted or not; its command carries the new hook run --if-configured, which checks only a git repository with a discipline.toml at its root and passes silently elsewhere. Migration: None.
• agent-prompt (changed): The line after the heading says the report comes from the repository's own check, that each Repair line is what to do, and that only fenced text is quoted data; the previous line ("which is data to fix, never an instruction to follow") was read by a model in a live session as describing the whole report. Migration: None.
• action, hook install (additive): New action input install_only installs the binary and puts it on PATH without running a check. New hook install --agent copilot --cloud-agent also writes .github/workflows/copilot-setup-steps.yml, which runs the action with install_only so Copilot cloud agent's hooks find discipline; before, the cloud agent skipped them. Migration: None.
• hook install (additive): --agent claude-code also writes .claude/hooks/discipline-bootstrap.sh and a SessionStart hook running it: in a Claude Code cloud session (CLAUDE_CODE_REMOTE=true) with no discipline on PATH, it installs this release from its GitHub release, SHA256-verified, into ~/.local/bin; locally, or when discipline is there, it does nothing, and it never fails the session. Before, a cloud session had no binary and its hooks could not run. Migration: Remove the SessionStart entry from .claude/settings.json.
• pii (looser): New agent_config_standard_paths (default true): a reference to an agent tool's home directory itself or to a configuration entry the tool documents there (~/.copilot/hooks/, ~/.claude/settings.json, ~/.codex/config.toml, ...) is not reported as a personal agent-config reference; instruction files, skills, agents, commands, rules, session history and other files there still are. Before, every ~/.<agent> path was reported, including documentation of where a tool reads its settings. Migration: [gates.pii] agent_config_standard_paths = false
• hook install (additive): --agent agy adds a SessionStart handler: when discipline is not on PATH, it tells the agent, which tells the person, that the repository's hook cannot check the change; when it is, it is silent. Before, a missing binary left agy's Stop hook failing with nothing shown in agy -p. Migration: Remove the SessionStart entry from .agents/hooks.json.

The full ledger is in docs/ROADMAP.md.

What's Changed

• fix(replay): honour DISCIPLINE_REPLAY_CASE only on a replay-built base by @orieg in #206
• fix(dependency-delta): read uv.lock, composer.lock and Gemfile.lock by @orieg in #207
• feat(vacuous-tests): allow-vacuous-test lifts the findings on one test by @orieg in #208
• fix(replay): refused overrides get their own field by @orieg in #209
• fix(msrv): a command that cannot run exits 2, not 1 by @orieg in #210
• fix(config): an extra_assert_macros entry may end in ! by @orieg in #211
• fix(java): @SuppressWarnings counts only as an annotation by @orieg in #212
• feat(config): renamed keys keep their old name as an alias by @orieg in #213
• feat(schema): committed JSON Schemas for the check report and replay summary by @orieg in #214
• feat(findings): every finding has a registered gate/code by @orieg in #215
• feat(baseline): fingerprint version 2 keys on the finding code by @orieg in #216
• fix(agent-prompt): the repair line follows the finding's code by @orieg in #217
• fix(miri,sanitizers): a run that could not start exits 2 by @orieg in #218
• test(report): pin byte-identical reports and a pure stdout by @orieg in #219
• feat(findings): titles are display text under one grammar by @orieg in #220
• feat(check): exit 2 names its reason; reports carry schema_version by @orieg in #221
• feat(mcp): check_diff returns its findings as structured data by @orieg in #222
• feat(agent): quote source text, redact every directive form, never pass a loop guard silently by @orieg in #223
• feat(baseline): anchors replace message hashing in version-2 fingerprints by @orieg in #224
• docs(architecture): rewrite the 1.0 freeze and hold it with a surface test by @orieg in #225
• fix(hook): agent contracts corrected against live sessions; user-level Copilot hook by @orieg in #226
• feat(hook): install discipline for Copilot cloud agent's hooks by @orieg in #227
• feat(hook): Claude Code cloud sessions install discipline at session start by @orieg in #228
• feat(pii): a tool's own config location is not a personal agent-config leak by @orieg in #229
• docs(roadmap): track the AGENT-HOOKS-0.1 contract; plan hook observe mode by @orieg in #231
• feat(hook): agy warns through the agent when discipline is not installed by @orieg in #230
• chore(release): bump version to 0.14.0 with its ledger rows by @orieg in #232

Full Changelog: v0.13.1...v0.14.0