Skip to content

CHANGELOG.md

Latest commit

 

History

History
465 lines (354 loc) · 49.8 KB

CHANGELOG.md

File metadata and controls

465 lines (354 loc) · 49.8 KB

Changelog

[2.20.1] - 2026-04-28

Fixed

  • Stop hook blocking legitimate stoppages when background agents / scheduled wakeups are in flight: The stop-hook validator (src/pacemaker/prompts/stop/stop_hook_validator_prompt.md) had a "BACKGROUND TASKS ARE REAL" exception, but its keyword list was too narrow ("running in background", "background job", "I'll be notified when it completes", "waiting for background", "run_in_background", "background process") and missed the way Claude actually phrases async delegation in practice — phrases like "two agents in flight", "will report consolidated findings when both return", and "agent returns" slipped past detection. Worse, the exception was framed as a single bullet inside the "agent still running fallacy" section, so when the LLM validator processed the conversation top-to-bottom and reached the strict "ALL DEVELOPMENT WORK REQUIRES E2E EVIDENCE" block (or saw "show-stopper" / "production blocking" framing from earlier user messages), it would override the exception and BLOCK the stop, trapping Claude in an awake state while delegated subagents were still working. This wasted context, burned tokens, and prevented the legitimate asynchronous workflow Claude Code is designed to support
  • Stop hook now recognizes ScheduleWakeup as a parallel async mechanism: The validator previously knew only about run_in_background=True and treated scheduled wakeups as "Claude avoiding completion". Both mechanisms automatically re-awaken Claude when ready, so both must trigger the APPROVED path. Added explicit recognition of scheduled-wakeup signals: "scheduled wakeup", "scheduled to resume", "ScheduleWakeup", "wakeup at", "next firing", "I'll resume in", "scheduled fire-time", "the loop will continue at", "scheduled re-entry", "drives the loop closure", "explicit wakeup", and similar
  • Stop hook now has an explicit precedence clause: The expanded "BACKGROUND TASKS / SCHEDULED WAKEUPS ARE REAL (HIGHEST PRECEDENCE)" section now states explicitly that this exception OVERRIDES "show-stopper / blocker / critical / production blocking" framing, the E2E-evidence requirement (the in-flight agents may BE doing the E2E work — demanding it from the orchestrator before they return is impossible by definition), "analysis paralysis" suspicion (delegation IS the action being taken), user urgency phrases ("can't release", "must complete this session", "ASAP"), and apparent unfinished work in the LAST MESSAGE (it IS being worked on, asynchronously). Only a visible launch FAILURE (error / exception / "agent crashed" / "wakeup rejected") with no recovery in progress overrides the exception
  • Stop hook now has a precedence reminder at the top of the E2E section: A one-line reminder was added to the top of "ALL DEVELOPMENT WORK REQUIRES E2E EVIDENCE" pointing back to the background/wakeup exception, so the LLM validator does not lose the precedence rule when it reaches the strict E2E block. The "WHEN TO BLOCK" bullet about "agent/slash command still running" was also updated to reference both background tasks AND scheduled wakeups as the legitimate exceptions
  • Net behavior: When Claude's last message announces background-agent delegation or a scheduled wakeup (alone or in combination), the stop hook now correctly returns APPROVED and lets the runtime resume Claude when the async work completes, instead of forcing Claude to stay awake and burn context

[2.20.0] - 2026-04-18

Added

  • Memory Localization — git-tracked .claude-memory/ with auto-linking (story #65): New feature that makes Claude Code's per-project memory portable across clones by symlinking the central ~/.claude/projects/<encoded>/memory/ folder to a repo-local .claude-memory/ directory that developers commit to git. Three flows: (A) SessionStart auto-link — when .claude-memory/ exists at the git root, the hook replaces the central memory folder with a symlink to it; local always wins, central content is safely renamed to memory.bak_localize then deleted only after symlink succeeds, with rename-back rollback on OSError. (B) CLI seed pace-maker localize-memory — copies central memory contents into <repo>/.claude-memory/ then replaces central with symlink; refuses if .claude-memory/ already exists. (C) CLI unlink pace-maker memory-localization unlink — removes the symlink and copies .claude-memory/ contents back to central; leaves the repo folder in place. Config gate memory_localization_enabled (default true) togglable via pace-maker memory-localization on/off/status and displayed in pace-maker status. On all success states (linked_fresh, replaced_with_symlink, relinked, already_linked, raced_but_ok), the hook injects a nudge via hookSpecificOutput.additionalContext telling Claude the memory is git-tracked
  • Path discovery without re-implementing encoding: Flow A derives the encoded central folder from SessionStart hook_data.transcript_path (Path(transcript_path).parent); Flow B/C scan ~/.claude/projects/*/*.jsonl matching the project's cwd. No re-implementation of Claude Code's path-encoding scheme
  • Safety invariants: assert_safe_to_destroy(path) requires path under CENTRAL_BASE and path.name == "memory" before any rmtree; replace_with_symlink_atomic renames to .bak_localize, symlinks, rmtree's the backup — on OSError the rename is reversed; _is_under uses canonicalize-parent-only boundary check so symlink leaves do not escape
  • Optimistic race handling: On FileExistsError during symlink creation, re-classify; if now correct_symlink return raced_but_ok (two concurrent SessionStart hooks for the same project race safely to convergent state)
  • PACEMAKER_CENTRAL_BASE test-mode enforcement: Mirrors PACEMAKER_SESSION_REGISTRY_PATH precedent — core.py raises RuntimeError when PACEMAKER_TEST_MODE=1 and the env var is unset, preventing tests from accidentally polluting real ~/.claude/projects/. CENTRAL_BASE resolved dynamically per access via module-level __getattr__ so per-test monkeypatch.setenv changes take effect
  • Classification states: missing, correct_symlink, wrong_symlink, regular_folder, permission_denied, unknown — explicit handling for each, unknown states logged and skipped rather than destroyed
  • 33 unit tests across tests/test_memory_localization_classification.py, test_memory_localization_linking.py, test_memory_localization_seed_restore.py — including deterministic race-condition test (raced_but_ok via monkeypatched os.symlink that creates-then-raises-FileExistsError) and OSError rollback test (monkeypatched os.symlink raising OSError, assert backup renamed back and content intact)
  • E2E validation against real Claude Code binary: Headless claude -p in a tmp git repo with .claude-memory/ pre-seeded with synthetic unique facts; verified Claude's internal memory tooling accesses the central symlinked path (~/.claude/projects/.../memory/) which transparently resolves through the symlink to workspace .claude-memory/ files
  • Project CLAUDE.md Memory Localization section: Architecture notes, Flow A/B/C descriptions, CLI commands, config gate, key-files map, test-sandbox pattern
  • pace-maker status memory-localization line: Displays Memory Localization: ENABLED/DISABLED between Danger Bash and Model Preference

[2.19.3] - 2026-04-18

Changed

  • SessionStart intent validation guidance — "Senior Coding Nanny" framing: Added a NAMING directive to src/pacemaker/prompts/session_start/intent_validation_guidance.md instructing Claude to refer to validation blockages as coming from the "Senior Coding Nanny" rather than "the hook system", "the hooks blocked me", or similar passive phrasing. The underlying blocker is a reviewer LLM agent making a judgment call on whether the declared INTENT matches the diff — not a dumb regex or a mechanical hook. Renaming the actor surfaces that reality to the user and stops Claude from blaming infrastructure for decisions made by a reasoning agent

[2.19.2] - 2026-04-16

Fixed

  • Stop hook death spiral on valid E2E evidence: When the stop hook rejected an E2E evidence table, it injected the rejection as an isMeta: true user message. The assistant responded with a short "I'm waiting" message. On the next stop attempt, the backwards-walk context evaluator saw the short message as the "last message" — not the E2E table. Each rejection generated a new short response, pushing the real evidence further down and creating an infinite rejection loop. Fix: build_stop_hook_context() in transcript_reader.py now filters isMeta entries and short (<200 char) assistant responses that immediately follow META messages. Real substantive content (>200 chars) after META is preserved. 5 new tests in test_stop_hook_meta_filtering.py
  • Overly rigid E2E evidence format check: The stop hook validator prompt demanded exact column headers (| # | AC | Test Description | How Performed | Real System / Data | Observed Result |) for FORMAT B. Real E2E evidence tables from subagents used similar but different headers like | # | Test | Command | Captured Output | Result | — containing the same substantive information (test name, execution method, real captured output, pass/fail) but with different column names. Added FORMAT C to stop_hook_validator_prompt.md accepting any markdown table with 3+ rows of real commands, real captured output, and per-row results regardless of exact column headers

[2.19.1] - 2026-04-15

Added

  • Cross-Session Awareness Registry (story #64): New feature that prevents "rogue agent" hallucinations when multiple Claude Code sessions work concurrently on the same repository. Shared SQLite registry at ~/.claude-pace-maker/session_registry.db (WAL mode, 2s busy_timeout) tracks live sessions keyed on canonical workspace root resolved via git rev-parse --show-toplevel (fallback os.path.realpath(os.getcwd())). Sibling discovery fires at SessionStart (startup/resume), at SubagentStart via hookSpecificOutput.additionalContext, periodically every 5 PreToolUse calls per agent_id, and before destructive Bash commands (injected into Stage 2 danger_bash validation context). Heartbeats on every hook invocation (SessionStart, SubagentStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop, SubagentStop, Notification, PreCompact, SessionEnd) with 20-minute stale-row purge. Config gate cross_session_awareness_enabled (default true) with CLI toggle pace-maker cross-session-awareness on/off. Best-effort graceful unregistration on SessionEnd. Fail-open semantics: all registry operations wrapped in broad exception handlers, hook path never crashes on registry errors
  • pace-maker sessions list CLI command: Renders active sibling sessions with workspace, PID, start time, last seen. Stale-row filter (>20 min) applied at render time
  • pace-maker cross-session-awareness on/off CLI command: Toggles the feature flag in ~/.claude-pace-maker/config.json
  • Namespace-safe state access: All non-SessionStart CSA handlers use defensive state.get("cross_session_awareness") pattern so mid-session config enable/disable does not crash subsequent hooks
  • PACEMAKER_SESSION_REGISTRY_PATH test-mode enforcement: db.py raises RuntimeError when PACEMAKER_TEST_MODE=1 is set but the env var is unset, preventing tests from accidentally polluting the live registry
  • Multi-process concurrency test: test_three_concurrent_heartbeats_succeed uses real multiprocessing.Process (fork-based) to verify WAL + busy_timeout handles 3 concurrent writers without crash or corruption
  • Project CLAUDE.md Cross-Session Awareness Registry section: Architecture notes, key file map, CLI reference, config gate, and state schema

Fixed

  • CSA state pollution across concurrent workspaces (critical bug in 2.19.0 story before ship): The original story design assumed ~/.claude-pace-maker/state.json was per-session, but it is a single global file shared across all live Claude Code sessions. The cross_session_awareness namespace was a flat dict, so when a pace-maker session wrote {"workspace_root": "/.../claude-pace-maker", ...}, it overwrote a concurrent code-indexer session's cached workspace_root. The code-indexer session's next hook then queried siblings in the pace-maker workspace and received pace-maker sibling sessions, leaking across unrelated repositories. Fix: scope the entire CSA namespace by session_id so each session reads and writes only its own sub-dict state["cross_session_awareness"][<session_id>] = {workspace_root, seen_agent_ids, tool_use_counter}. Added 6 session-scoping regression tests and a legacy-flat-format migration path. Verified via live outside-in reproduction: a session launched in /tmp sees zero pace-maker siblings, a session launched in the pace-maker workspace sees only pace-maker siblings (never the /tmp one). on_session_end now GCs the session's sub-dict to prevent unbounded growth of state.json. See src/pacemaker/session_registry/_csa.py and tests/test_session_registry_csa_session_scoping.py
  • Pre-tool hook _csa_result UnboundLocalError: _csa_result was declared deep inside the run_pre_tool_hook try block, so any exception before that line (malformed stdin, JSON decode errors, transcript path resolution) caused the outer except Exception handler to crash with UnboundLocalError when calling _merge_csa_reminder(..., _csa_result). Fix: hoisted the initializer to the first statement inside the try block. Added test_fails_open_on_invalid_stdin_json_no_unbound_local_error regression test that feeds invalid JSON to the hook and asserts {"continue": True} fail-open
  • build_periodic_reminder UnicodeEncodeError (surrogate pair): _REMINDER_HEADER used \ud83d\udd14 (UTF-16 surrogate pair escape for 🔔) which Python holds in memory but crashes on UTF-8 encoding (e.g., safe_print to stdout). Would have broken AC3 (periodic reminder every 5th PreToolUse) in production. Fix: changed to \U0001F514 (single-codepoint escape). Added three encoding-safety tests for all nudge builders that call .encode("utf-8") on the builder output
  • SubagentStart double hookSpecificOutput emission: The refactor that extracted _emit_subagent_additional_context() helper called it twice (once for intent validation guidance, once for CSA banner), emitting two JSON objects concatenated on stdout. Claude Code's hook parser expects exactly one JSON object per SubagentStart invocation, so one message would be silently dropped. Fix: accumulate both context strings into a local list and emit once at the end of the handler with blank-line separator. Added integration test that verifies exactly one JSON object is emitted containing both context texts
  • AC3 periodic reminder dropped on Write/Edit return paths: _csa_result.get("periodic_reminder") was only merged into hookSpecificOutput on the tool_name not in ["Write", "Edit"] path. The Write/Edit branch had 8 return sites that all skipped the merge (no file_path, master disabled, feature disabled, non-source file, unknown tool, approved, blocked, exception handler). Fix: added _merge_csa_reminder(response, csa_result) helper and wrapped every Write/Edit return with it. Added 10 parameterized wiring tests covering all 8 return paths
  • Regressions in existing test files from hook.py CSA wiring: Updated mock signatures in test_two_stage_validation.py, test_stage2_classification_parsing.py, test_pre_tool_hook.py, test_clean_code_rules_e2e.py, test_langfuse_orchestrator_incremental_spans.py, and test_secrets_e2e.py to match the new reviewer-identity tuple return shape and state isolation fixtures. No assertion weakening — only mock-shape corrections
  • tests/conftest.py registry isolation: Added autouse fixture that sets PACEMAKER_SESSION_REGISTRY_PATH to a per-test tmp path under the fake HOME, preventing pre-existing tests that happen to trigger a CSA path from polluting the live registry

Changed

  • Version bump 2.18.2 → 2.19.1: Minor bump for the cross-session awareness feature (planned as 2.19.0) plus a patch bump for the pre-ship state-pollution fix documented above
  • scripts/pace-maker: Restored to bash shim (had been accidentally converted to Python, breaking the plugin install path's 5 CLI wrapper tests)
  • src/pacemaker/constants.py: Added DEFAULT_DANGER_RULES_PATH constant so _csa.on_pre_tool_use and the main danger_bash block share a single source of truth instead of duplicating the os.path.expanduser("~/.claude-pace-maker/danger_bash_rules.yaml") literal

[2.18.2] - 2026-04-14

Fixed

  • Competitive expression CLI regex: pace-maker hook-model gpt-5.4+gemini-flash->haiku was silently dropped by parse_command() because the competitive-expression regex character class [a-z0-9-]+ did not include the literal .. The canonical form advertised in help text therefore never reached the hook-model handler — only the legacy alias gpt-5+... worked by accident. Character class widened to [a-z0-9.\-]+; semantic validation is still delegated to parse_competitive(). New tests added for both gpt-5.4 and legacy gpt-5 competitive parse paths via parse_command

Changed

  • Centralized model alias map: Extracted KNOWN_MODELS and SHORT_ALIASES into a new leaf module src/pacemaker/inference/model_aliases.py imported by both competitive.py and codex_provider.py. Eliminates the hardcoded inline gpt-5 → gpt-5.4 check previously duplicated in codex_provider.query() (Messi rule 4, anti-duplication). Single source of truth; no circular import between sibling providers
  • Stale docstrings/comments updated: gpt-5 references in docstrings and comments across code_reviewer.py, constants.py, inference/provider.py, inference/registry.py, intent_validator.py, and user_commands.py replaced with canonical gpt-5.4 (backward compat note retained where helpful)

[2.18.1] - 2026-04-14

Fixed

  • gpt-5.4 Codex CLI model identifier: gpt-5.4 is now the canonical model token for Codex CLI (ChatGPT account subscription). The previous token gpt-5 is retained as a backward-compatible alias that transparently maps to gpt-5.4 — existing configs and competitive expressions using gpt-5 continue to work without change. The Codex CLI subprocess always receives -m gpt-5.4, fixing the ERROR: The 'gpt-5' model is not supported failure that caused the competitive reviewer pipeline to fall back to SDK solo

[2.18.0] - 2026-04-08

Added

  • Stop hook E2E enforcement protocol: Replaced narrow story/epic-only E2E gate with a 3-step protocol covering ALL development work (story/epic, bug fixes, feature additions, behavioral refactors). Step 1 detects development work from session context; Step 2 checks for valid exits (EXIT A: specific not-applicable justification validated by LLM; EXIT B: verbatim user waiver quote, identical rules to TDD override); Step 3 requires E2E evidence in accepted format before allowing stop
  • E2E evidence formats: Stop hook now accepts FORMAT A (E2E TEST COMPLETION REPORT from e2e-test-heuristic.md standard with CHANGED CODE COVERAGE + REGRESSION COVERAGE + OVERALL VERDICT) or FORMAT B (E2E Evidence Table with #, AC, Test Description, How Performed, Real System / Data, Observed Result columns); running pytest or claiming "tests pass" explicitly rejected
  • Stop hook exit valve counter: New consecutive_stop_blocks field in session state prevents infinite block loops — increments on each BLOCKED stop, resets to 0 on any tool use (PostToolUse handler), releases on 5th consecutive block (STOP_EXIT_VALVE_THRESHOLD = 4); "EV" activity event recorded on activation
  • STOP_EXIT_VALVE_THRESHOLD module constant: Replaces local variable, imported by tests to eliminate threshold duplication across files
  • Minimum 3 E2E tests per acceptance criterion: e2e-test-heuristic.md now mandates happy path + edge case + error/boundary test per AC, enforced by stop hook's minimum row count requirement
  • All-dev-work E2E scope in standards: testing-quality-standards.md updated to state E2E is mandatory for all development (not just story/epic), stop hook enforcement noted, completion criteria strengthened to require accepted format with 3+ rows and real captured output

Fixed

  • test_blockage_telemetry.py stale count: Pre-existing test failure fixed — category count assertion corrected from 6 to 7 to reflect intent_validation_dangerbash category added with Danger Bash feature

[2.16.0] - 2026-04-07

Added

  • 5 new default clean code rules: type-safety-erosion (TypeScript any, Java raw types, C# dynamic, Go interface{}, Kotlin !!), ignored-error-return (Go _, err discard, Java/Kotlin discarded Result), unhandled-async (missing await, async void, fire-and-forget Promises), hardcoded-config (env-specific URLs/ports/paths distinct from secrets and magic numbers), unsafe-string-interpolation (command injection and XSS via string interpolation, extends sql-injection coverage)

Changed

  • Replaced mutable-defaults with unsafe-defaults: Generalizes Python-only mutable default argument rule to cover JS/TS module-level mutable defaults and Java/Kotlin static mutable fields used as defaults
  • Patched 5 rule descriptions for multi-language coverage: exception-handling (adds Go val, _ := fn() idiom), resource-leak (adds Java try-with-resources, C# using, Go defer), path-traversal (adds Java/Go/Node examples), concurrency-hazard (generalizes from Python global/module-level to Go/Java/C#/JS), hidden-magic (adds JS eval, Java/Kotlin reflection, Go reflect, C# dynamic)
  • Default rule count: 20 → 25

[2.15.0] - 2026-04-06

Added

  • Competitive multi-model review pipeline (#63): New hook_model expression syntax m1+m2[+m3]->synthesizer dispatches 2–3 reviewers in parallel and synthesizes results; supported models: auto, sonnet, opus, haiku, gpt-5, gemini-flash, gemini-pro
  • inference/competitive.py (#63): Core module with parse_competitive() (validates expression), run_competitive() (parallel dispatch + synthesis), _dispatch_reviewers() (ThreadPoolExecutor with per-reviewer timeout), _synthesize() (synthesis phase with bounded timeout)
  • Synthesis formatter prompt (#63): Synthesizer acts as formatter not judge — mechanically consolidates verdicts: any BLOCK from any reviewer → BLOCKED; all APPROVED → APPROVED; does not add or remove concerns
  • Reviewer verdict logging at DEBUG (#63): First 300 chars of each reviewer's raw response logged via log_debug() for synthesis quality evaluation (MAX_REVIEW_LOG_CHARS = 300)
  • COMPETITIVE REVIEW MODE: section in pace-maker help (#63): Documents expression syntax, model aliases, supported reviewers, failure modes, and examples
  • Named constants for timeouts and sizing (#63): REVIEWER_WAIT_TIMEOUT_SEC = 60, SYNTHESIS_TIMEOUT_SEC = 30, MIN_REVIEWERS = 2, MAX_REVIEWERS = 3 replace magic numbers throughout competitive.py

Changed

  • pace-maker status competitive display (#63): Shows full expression (e.g. opus+gpt-5->haiku) in ANSI blue instead of "competitive" + separate "reviewers:" breakdown line
  • PreToolUse hook timeout (#63): Bumped from 60s to 120s in ~/.claude/settings.json to accommodate 2-phase pipeline (reviewer dispatch + synthesis) without mid-execution kill
  • Synthesis timeout bounded (#63): _synthesize() wraps synthesizer call in future.result(timeout=SYNTHESIS_TIMEOUT_SEC) — timeout → first-survivor-wins; avoids indefinite block

Fixed

  • Individual phase timeouts (#63): Per-reviewer timeout (futures_wait(timeout=60s)) and synthesis timeout (future.result(timeout=30s)) replace a single unbounded outer timeout, ensuring partial results are always preserved on slow reviewers

[2.14.0] - 2026-04-05

Added

  • Gemini CLI inference provider (#54): Two new hook model options — gemini-flash (→ gemini-2.5-flash) and gemini-pro (→ gemini-2.5-pro) — using Google Gemini CLI subprocess with stdin prompt injection (avoids process-list exposure)
  • Short CLI aliases (#54): pace-maker hook-model gem-flash and pace-maker hook-model gem-pro accepted as aliases, stored canonically as gemini-flash/gemini-pro
  • Reviewer identity tracking for Gemini (#54): [REVIEWER:gem-flash] and [REVIEWER:gem-pro] tags in governance event feedback_text; displayed as [Gem] (cyan) in claude-usage monitor
  • GEM-FLASH / GEM-PRO in pace-maker status (#54): Short display names shown instead of raw model strings
  • Fallback chain for Gemini (#54): Gemini failure falls back to auto (Anthropic) then fail-open; same chain as Codex provider
  • GeminiProvider in inference/ package (#54): GeminiProvider exported from inference/__init__.py; inference/ subdirectory now deployed by install.sh

[2.13.0] - 2026-04-05

Added

  • Danger bash CLI CRUD commands (#62): Full command-line management matching clean-code rules UX — pace-maker danger-bash list/add/remove/restore/modify/on/off
  • danger_bash_cli.py helper module (#62): Thin CLI dispatcher with argument parsing, regex validation on add (re.compile before write), category validation, and clear error messages
  • CRUD functions in danger_bash_rules.py (#62): add_rule(), remove_rule(), restore_rule(), modify_rule(), _write_config() (atomic write with directory creation), format_rules_for_display() (source tags, pattern truncation, stable sort)
  • pace-maker status shows Danger Bash (#62): Displays ENABLED/DISABLED state alongside Intent Validation and TDD
  • Help text updates (#62): All danger-bash commands in pace-maker help with DANGER BASH RULES documentation section
  • Same-ID warning in loader (#62): load_rules() logs warning when custom rule ID matches a default (skips custom, uses default)
  • 54 CRUD unit tests (#62): Comprehensive coverage for add/remove/restore/modify, validation, atomic write, display formatting

Changed

  • CLI title (#62): Updated from "Credit-Aware Adaptive Throttling" to "Credit-Aware Adaptive Throttling... and much more"

[2.12.0] - 2026-04-05

Added

  • Danger Bash intent validation (#58): Two-phase validation for dangerous Bash commands — Phase 1 regex gate fast-rejects commands matching danger rules when no INTENT: declaration is present (no LLM call); Phase 2 LLM validates intent-to-command alignment when INTENT: is present
  • 55 default danger bash regex rules (#58): danger_bash_rules_default.yaml with 25 Work Destruction (WD) rules and 30 System Destruction (SD) rules covering git destructive operations, file deletion, process killing, and infrastructure commands
  • danger_bash_rules.py module (#58): Loads bundled defaults, merges with user customizations from ~/.claude-pace-maker/danger_bash_rules.yaml, pre-compiles regex patterns at load time, and matches commands against compiled rules (same merge strategy as clean code rules)
  • intent_validation_dangerbash blockage category (#58): New blockage category with "Danger Bash" label for telemetry and blockage stats
  • 483 parametrized regex tests (#58): Comprehensive test coverage for all 55 rules including loader, match (WD and SD categories), performance, and edge cases
  • Reviewer identity tracking (#60): resolve_and_call_with_reviewer() returns (response, reviewer_name) tuple; reviewer identity threaded through intent_validator to hook to blockage_events details JSON; governance event feedback_text prefixed with [REVIEWER:xxx] tag
  • _refresh_codex_usage() on fallback (#60): Codex usage stats refreshed when Codex inference fails and falls back to auto provider

Changed

  • PreToolUse hook tool matching (#58): Expanded from Write|Edit to include Bash tool for danger bash validation (existing Write/Edit validation unchanged)
  • Installer updated (#58): Deploys danger_bash_rules_default.yaml config file alongside existing YAML configs

Fixed

  • Codex PAYG billing crash (#59): _parse_last_token_count() no longer crashes when Codex CLI returns null primary/secondary fields for PAYG billing plans
  • limit_id column in codex_usage table (#59): Added with idempotent ALTER TABLE migration; wired migration call in hook.py SubagentStop handler

[2.11.0] - 2026-04-04

Added

  • Clean code rules merge strategy (#55): YAML config now stores only custom rules and deletion markers, not defaults. load_rules() merges at runtime: defaults minus deleted, overrides applied at position, custom appended
  • _load_custom_config(): Reads YAML returning {rules, deleted_rules} with validation for missing files, empty YAML, invalid types, and malformed entries
  • _write_config(): Atomic write via temp file + os.replace() with orphan cleanup on failure (replaces _write_rules)
  • _migrate_snapshot(): One-time migration strips old snapshot copies of defaults, preserving only genuine overrides and custom rules
  • get_rules_metadata(): Returns [{id, source}] with source as "default", "override", or "custom" for display tagging
  • Source tags in CLI: pace-maker clean-code list now shows [default], [override], [custom] tags per rule
  • 4 new security rules: credential-construction, path-traversal, resource-leak, concurrency-hazard
  • 58 unit tests + 7 integration tests covering all 24 acceptance criteria

Changed

  • Default rules refactored from 25 to 20: Merged redundant rules (bare-except + swallowed-exceptions → exception-handling, 4 size rules → blob-size, 2 mock rules → mock-abuse, undeclared-fallbacks → silent-degradation), removed context-dependent rules (missing-comments, missing-invariants, excessive-indirection, commented-code), sharpened descriptions for micro-review detectability
  • add_rule(): Only writes custom rule to YAML, clears deletion marker if re-adding
  • remove_rule(): Deletion marker for defaults, direct removal for custom, removes override AND suppresses default
  • modify_rule(): Override copy for defaults, in-place for custom, strips id from updates
  • format_rules_for_display(): Optional config_path parameter for source tagging

[2.10.0] - 2026-04-04

Added

  • Codex GPT-5 usage tracking (#57): New src/pacemaker/codex_usage.py module extracts subscription rate limits from Codex session JSONL files (~/.codex/sessions/YYYY/MM/DD/*.jsonl) after each codex subagent run
  • codex_usage SQLite table: Single deterministic record (id=1, INSERT OR REPLACE) in usage.db storing primary_used_pct, secondary_used_pct, primary_resets_at, secondary_resets_at, plan_type, and timestamp — same pattern as api_cache and backoff_state
  • SubagentStop hook integration: Automatically captures codex rate limits when hook_model contains "gpt" or "codex"; wrapped in isolated try/except to never break existing hook logic
  • 18 new unit tests: tests/test_codex_usage.py covering JSONL parser (valid data, missing fields, malformed JSON, no sessions, historical fallback, mtime-based file selection) and SQLite read/write (single record, overwrite, DB errors)

[2.7.0] - 2026-03-30

Added

  • Multi-model inference provider abstraction (#53): New src/pacemaker/inference/ package enabling non-Anthropic models for hook validation
  • InferenceProvider abstract base class with ProviderError exception for provider-agnostic inference
  • AnthropicProvider: Wraps existing claude_agent_sdk calls with automatic sonnet↔opus fallback on usage limits
  • CodexProvider: Shells out to OpenAI Codex CLI (codex exec) for GPT-5 inference with 120s timeout
  • Provider registry: get_provider() factory, resolve_model_for_call() for per-call-site defaults, resolve_and_call() orchestrator with cross-vendor fallback chain (selected provider → auto/Anthropic → fail-open)
  • hook_model config setting: New setting separate from preferred_subagent_model — controls which model pace-maker uses for its own inference (intent validation, code review, stop hook)
  • pace-maker hook-model CLI command: Set hook inference model via pace-maker hook-model [auto|sonnet|opus|gpt-5] (both real CLI and pseudo-CLI)
  • Hook model in status/help: hook_model value displayed in pace-maker status and pace-maker --help
  • 29 new unit tests: test_inference_provider.py covering provider contracts, CodexProvider subprocess handling, registry routing, resolve_and_call fallback chain, and config defaults

Changed

  • Refactored all 5 inference call sites to use resolve_and_call() provider abstraction:
    • intent_validator.py: stop hook validation, intent declaration check, stage 1 validation, stage 2 unified validation
    • code_reviewer.py: code review validation
  • Sync wrappers bypass async functions: Refactored sync wrappers (call_sdk_validation, _call_sdk_intent_validation, _call_stage1_validation, _call_stage2_validation) to call resolve_and_call() directly, avoiding nested event loop errors
  • claude-usage display: Relabeled "Model:" to "Subagent:" for preferred_subagent_model, added new "Hook Model:" line showing hook_model value

[2.6.0] - 2026-03-30

Added

  • Governance event feed (#52): New governance_events SQLite table records intent validation rejections (IV), TDD enforcement failures (TD), and clean code violations (CC) with full feedback text, project name, and session ID
  • record_governance_event(): Writes governance events to SQLite with WAL mode and 5s timeout, returns True/False (never raises)
  • cleanup_old_governance_events(): Purges events older than 24h, called from SessionStart hook
  • Hook integration: Governance events recorded at all 3 rejection paths alongside existing record_blockage() calls
  • 13 new unit tests: test_governance_events.py covering table creation, IV/TD/CC insertion, success/error returns, concurrent writes, and cleanup

[2.5.0] - 2026-03-26

Added

  • 8 new Messi-derived clean code rules: mock-in-e2e, over-engineering, code-duplication, orphan-code, unbounded-loops, missing-invariants, excessive-indirection, hidden-magic — total rules now 25 (was 17)
  • 16 new unit tests for Messi rules including rule ID uniqueness guard

Changed

  • Enhanced 4 existing rule descriptions: undeclared-fallbacks (covers alternative code paths, "just in case" logic), swallowed-exceptions (covers unchecked return values, explicit handling strategies), over-mocking (focused on unit test scope), large-files (specific thresholds: Scripts >200, Classes >300, Modules >500)

[2.4.0] - 2026-03-24

Fixed

  • Stage 2 rejection categorization non-deterministic (#51): Replaced fragile keyword matching (clean_code_keywords list) with deterministic CLASSIFICATION: structured output parsing from stage 2 code review responses. All stage 2 rejections now correctly categorize as intent_validation_cleancode instead of being silently miscategorized as intent_validation
  • Stage 1 intent validation false rejections: Softened the stage 1 declaration check prompt to accept INTENT: marker anywhere in the message, not just at the start. Added explicit guidance and valid examples showing mid-message intent declarations

[2.3.0] - 2026-03-10

Removed

  • Legacy pacing algorithm: Removed calculate_logarithmic_target() and calculate_delay() from calculator module — adaptive algorithm now runs unconditionally
  • use_adaptive parameter: Removed from calculate_pacing_decision() signature — no longer needed since adaptive is the only algorithm
  • algorithm key from pacing decisions: Return dict no longer contains "algorithm": "adaptive" or "algorithm": "legacy" — the Algorithm display line in claude-usage-reporting monitor is removed

Added

  • 20 acceptance tests: test_legacy_removal.py validates all legacy code paths are removed and adaptive runs unconditionally

[2.2.1] - 2026-03-10

Fixed

  • Fallback mode 5-hour window rollover (#46): _get_synthetic_snapshot() now correctly filters accumulated costs by the rollover boundary timestamp, reporting only post-rollover usage instead of inflated values from the prior window
  • Fallback rollover persistence independence (#46): _get_synthetic_snapshot() now persists rollover state independently, eliminating the ordering dependency on get_reset_windows() being called first
  • 7-day fallback rollover (#46): Same rollover fix applied to the 7-day window path, ensuring consistent behavior across both window types

[2.2.0] - 2026-03-09

Added

  • pace-maker install claude-usage-monitor command: New subcommand installs the claude-usage-monitor tool from its repository (Story #45)
  • HTTPS-first git auth detection: detect_git_auth() now tries plain HTTPS before falling back to SSH, improving compatibility with public repositories
  • Independent test runner: scripts/run_tests.sh runs each test file in its own pytest process with a 30-second timeout, eliminating SQLite WAL contention hangs during test suite execution

Fixed

  • install.sh CLI path detection: Corrected detection of the pace-maker CLI executable path during installation
  • Python 3.9 version checks: All version gating now correctly handles Python 3.9 (requirement lowered from >=3.10 to >=3.9)
  • --version flag verification: Installer now verifies the deployed CLI responds correctly to --version
  • SQLite WAL contention in tests: Resolved hangs caused by concurrent test files sharing WAL-mode databases in the same pytest process
  • Naive datetime bug: Fixed remaining naive datetime comparisons that caused crashes when mixing timezone-aware and naive objects

Changed

  • Minimum Python version: Lowered from >=3.10 to >=3.9 to broaden supported environments

[2.1.0] - 2026-03-08

Fixed

  • SA indicator color on stop: SubagentStop now correctly shows the blue indicator instead of an incorrect color
  • LF indicator gating: Langfuse (LF) indicator is only shown when a trace was actually pushed, not on every hook invocation
  • SM indicator gating: Secrets Masking (SM) indicator is only shown when secrets were actually masked in the trace, not on every sanitize call

[2.0.0] - 2026-03-08

Added

  • Global API poll coordination (#43): SQLite-backed singleton ensures only one session polls the Claude API at a time, eliminating redundant concurrent requests across parallel sessions
  • Activity events table: New activity_events SQLite table records hook invocations with timestamps for activity indicator computation
  • Hook instrumentation: All hook entry points record activity events for accurate timing and settings-awareness
  • Activity indicator help text: pace-maker help now documents all activity indicator symbols (PL, LF, SM, SS, SA) and their color meanings
  • PL indicator colors: Pacing (PL) indicator uses blue/yellow/red to reflect normal/warning/critical pacing state
  • Fallback coefficients in status: pace-maker status displays the active cost-to-utilization coefficients, including calibrated fallback values
  • COEFFICIENTS section in help: pace-maker help includes an explanation of how fallback mode coefficients work and how they are calibrated

Fixed

  • Naive/aware datetime mismatch: Replaced all datetime.now() and datetime.utcnow() calls in hook.py and supporting modules with datetime.now(timezone.utc), eliminating crashes in the pacing engine caused by mixing naive and aware datetime objects
  • Activity indicator timing: Corrected timing windows used to determine whether recent activity qualifies for indicator display
  • SS indicator gating: SessionStart (SS) indicator now fires only when new secrets are detected, not on every session start
  • SM indicator placement: Secrets Masking (SM) indicator fires in the orchestrator after sanitize_trace() completes, ensuring it reflects actual masking results
  • Global poll coordination code review findings (#43): Addressed follow-up issues identified during code review of the SQLite poll coordination feature

[1.19.0] - 2026-03-07

Added

  • Complete JSON-to-SQLite migration: All remaining state management (fallback state, backoff state, API cache, profile cache) migrated from JSON files to UsageModel SQLite tables, eliminating TOCTOU races between concurrent sessions
  • Test safety guard: conftest.py now sets PACEMAKER_TEST_MODE=1 to prevent test runs from polluting the production SQLite database

Removed

  • Dead api_backoff.py module: Deleted legacy JSON-based backoff implementation replaced entirely by UsageModel SQLite in v1.18.0
  • Dead fallback JSON code paths: Removed all remaining read/write paths that previously fell back to JSON files for state storage

[1.18.0] - 2026-03-07

Added

  • Resilient Fallback Mode: When the Claude API returns 429 errors, pace-maker enters fallback mode and synthesizes utilization estimates from accumulated token costs. Includes automatic state machine transitions (NORMAL → FALLBACK → NORMAL), per-session cost tracking in SQLite, and rollover-safe window projections
  • UsageModel — Single Source of Truth: New UsageModel class (src/pacemaker/usage_model.py) unifies all usage data access. Stateless between calls with all state in SQLite WAL mode. Both pace-maker hooks and claude-usage monitor read from the same source
  • Coefficient Calibration: When the API recovers after a fallback period, compares synthetic predictions against real API values. Auto-adjusts cost-to-utilization coefficients via weighted average, stored per tier (5x/20x) in calibrated_coefficients SQLite table
  • SQLite State Migration: Fallback state, API cache, backoff state, and profile cache moved from JSON files to SQLite tables (fallback_state_v2, api_cache, backoff_state, profile_cache). Eliminates TOCTOU races between concurrent sessions
  • Accumulated Cost Tracking: accumulated_costs table with INSERT-only concurrency-safe cost accumulation (no read-modify-write). Idempotent per-session deduplication prevents double-counting
  • Pressure Tests: 23 new tests covering 5h/7d cycle switching, coefficient calibration, rollover handling, and edge cases

Fixed

  • Rollover detection in synthetic mode: Fixed production bug where _project_window() returned five_rolled=False after get_reset_windows() persisted the projected window. Now checks persisted rollover_cost_5h/7d as primary indicator
  • Stale JSON fallback guard: is_fallback_active() no longer lets stale fallback_state.json override empty SQLite table during transition period
  • Tier-aware calibration: calibrate_on_recovery() stores under detected tier (e.g., "20x") from profile cache, not hardcoded "5x"
  • Per-project fallback lock: Scoped fallback lock file path to prevent cross-project interference

[1.17.0] - 2026-02-22

Added

  • Version bump TDD bypass: Stage 1 validation now skips TDD enforcement for version-bump-only changes to core path files, avoiding unnecessary test declarations for trivial version string updates

[1.11.0] - 2026-02-11

Fixed

  • Per-turn token counting: Generation observations now report tokens for the current turn only, not accumulated across the entire transcript. Fixes inflated cost reporting (e.g. $3.74 reported vs actual ~$0.02 per turn)
  • Subagent transcript path detection: Search new Claude Code 2.1.39+ nested directory structure (<session-id>/subagents/agent-*.jsonl) with backward compatibility for old flat structure

Added

  • Subagent generation observations: Subagent traces now include generation-create events with token usage and cost, matching main session behavior
  • Per-turn token counting tests: 3 tests for turn boundary detection and token scoping
  • Subagent generation tests: 4 tests for subagent cost tracking

[1.10.0] - 2026-02-10

Fixed

  • Langfuse trace pipeline: Fixed 12 bugs in trace/span lifecycle (deferred push, pending trace flush, subagent state corruption, BrokenPipeError protection)
  • Generation observation: Added generation-create event to stop hook for Langfuse totalCost computation (traces/spans alone don't compute cost)

Added

  • BrokenPipeError protection: All stdout writes use safe_print() to prevent hook crashes

[1.9.0] - 2026-02-09

Fixed

  • Intel prompt value formats: Enforce strict decimal/code formats to prevent text-based values that break the parser

[1.8.0] - 2026-02-08

Added

  • Prompt Intelligence (Intel): Per-prompt metadata telemetry (§ intel lines with frustration, specificity, task type, quality, iteration)
  • Intel Langfuse Integration: Parsed intel attached to Langfuse traces as intel_* metadata keys for dashboard filtering
  • Intel Guidance Prompt: Session-start injection of intel symbol vocabulary
  • Langfuse Provisioner E2E Tests: End-to-end test coverage for auto-provisioning

[1.7.0] - 2026-02-06

Added

  • Secrets Management: Sanitizes sensitive data (API keys, tokens, passwords) from Langfuse trace outputs before pushing

Fixed

  • Langfuse Tool Output Capture: Fixed tool output capture for accurate trace content

[1.6.0] - 2026-02-05

Added

  • Langfuse Auto-Provisioning: Automatic API key provisioning with configurable URL via pace-maker langfuse configure
  • Langfuse Status Display: Shows provisioning URL and connectivity in pace-maker langfuse status

[1.5.0] - 2026-02-04

Added

  • Daily Log Rotation: One log file per day (pace-maker-YYYY-MM-DD.log), 15 days retention
  • Enhanced Status Display: Shows versions, Langfuse connectivity, 24-hour error counts
  • Mypy Type Fixes: Resolved all type errors in langfuse modules

[1.4.1] - 2026-02-04

Added

  • Langfuse telemetry integration: Direct HTTP API integration for tracing Claude Code sessions
  • Blockage telemetry tracking: Track and report intent validation blockages via CLI stats
  • Model preference to status display: Shows preferred model in pace-maker status output
  • Stale data detection: Resilient pacing calculations handle stale/missing data gracefully

Fixed

  • Intent validation message count: Fixed to n=2 (minimum required because Claude Code writes text content and tool_use as separate transcript entries)
  • TDD blockage tracking: Proper tracking of TDD enforcement blockages
  • Stop hook tempo checker: More permissive handling of incomplete context

Changed

  • Package description: Updated to reflect full feature set (pacing, intent validation, TDD enforcement, Langfuse telemetry)

[1.5.0] - 2025-12-09

Added

  • Two-stage validation system: Separates fast declaration checking (Stage 1, ~2-4s) from comprehensive code review (Stage 2, ~10-15s)
  • Stage 1 (Fast Declaration Check): Uses Sonnet to validate intent declaration exists with all required components
  • Stage 2 (Comprehensive Code Review): Uses Opus (with Sonnet fallback) for deep code quality validation
  • TDD enforcement CLI: pace-maker tdd on|off command
  • Clean code rules CLI: pace-maker clean-code list|add|remove commands
  • Core paths CLI: pace-maker core-paths list|add|remove commands
  • Log level control: pace-maker loglevel 0-4 command (0=OFF, 1=ERROR, 2=WARNING, 3=INFO, 4=DEBUG)
  • Externalized clean code rules: ~/.claude-pace-maker/clean_code_rules.yaml
  • Externalized core paths: ~/.claude-pace-maker/core_paths.yaml
  • Centralized logging system: Configurable log levels across all modules

Changed

  • Message extraction: Combines last 2 messages to handle Claude Code's text/tool call splitting
  • Stage 1 model: Changed from Haiku to Sonnet for better intent detection
  • Stage 2 response format: Returns "APPROVED" text instead of empty string for pass
  • Model naming: All models use generic aliases (claude-sonnet-4-5, claude-opus-4-5, claude-haiku-4-5) that auto-update
  • Prompt organization: Reorganized into pre_tool_use/, common/, stop/, session_start/, user_commands/ directories
  • Exception handling: Validation now fails closed (blocks) on errors instead of failing open
  • Intent declaration requirement: Must be in SAME message as Write/Edit tool (not in prior messages)

Fixed

  • Installer: Now correctly copies Python modules to ~/.claude/hooks/pacemaker/
  • Model names: Fixed non-existent claude-haiku-4 model name
  • Thinking budget: Increased from 1000 to 1024 (API minimum requirement)
  • PyYAML dependency: Added for python3.11 compatibility

[1.4.0] - 2025-12-03

Added

  • Pre-tool intent validation: Claude must declare intent (FILE, CHANGES, GOAL) before code modifications
  • Light-TDD enforcement: Core code paths (src/, lib/, core/, source/, libraries/, kernel/) require test declarations or explicit user permission to skip
  • Clean code validation: Blocks 15 categories of violations including hardcoded secrets, SQL injection, bare except clauses, magic numbers, mutable defaults, over-mocked tests, and logic bugs
  • Code-intent alignment checks: Detects scope creep, missing functionality, and unauthorized deletions
  • 5-hour limit toggle: pace-maker 5-hour-limit on|off command
  • Intent validation toggle: pace-maker intent-validation on|off command
  • Test report: reports/pre_tool_validation_test_report.md documenting all validation behaviors

Changed

  • Pre-tool validator uses Opus as primary model with Sonnet fallback
  • Message context expanded to 5 messages (4 text-only + 1 full with tool parameters)

[1.3.1] - 2025-11-XX

Added

  • Subagent reminder system for main context delegation
  • Session lifecycle tracking (tempo) with AI validation

[1.3.0] - 2025-11-XX

Added

  • Weekend-aware throttling algorithm
  • 12-hour preload allowance for weekday starts
  • 95% safety buffer targeting

[1.2.0] - 2025-11-XX

Added

  • Dual window support (5-hour and 7-day limits)
  • Adaptive delay calculation

[1.1.0] - 2025-11-XX

Added

  • CLI interface (pace-maker command)
  • Configuration file support

[1.0.0] - 2025-11-XX

Added

  • Initial release
  • Basic credit throttling via PostToolUse hook
  • SQLite usage database