codebugs

A code-finding, requirements, and release tracker for AI assistants — one where a finding has an identity. SQLite-backed, exposed via an MCP server and a CLI.

Most trackers treat every report as a new row, so the second agent to notice the same bug files it again, and the queue fills with copies of one defect. codebugs treats a finding as a defect, and each report of it as one observation of that defect:

$ codebugs add -s high -c n_plus_one -f src/api.py -d "Query in loop at line 42" --new-category
Added: CB-1
$ codebugs add -s high -c n_plus_one -f src/api.py -d "Query in loop at line 42"
Bumped: CB-1 (occurrence 2)
$ codebugs update CB-1 --status fixed
Updated: CB-1 (status=fixed, severity=high)
$ codebugs add -s high -c n_plus_one -f src/api.py -d "Query in loop at line 42"
Reopened as regression: CB-1 (occurrence 3)

One card, three observations, and a regression recorded on the row it belongs to. Filing an observation again is normal and useful, not noise.

What makes this different

Deduplication is the point, not a side effect — and the rest of the design falls out of that one decision.

• Filing the same finding twice does not create two cards. The second report bumps the first: its occurrence count goes up, and its severity rises if this sighting was worse than the last. Severity only ever escalates under observation, so a card filed low and re-seen critical stops hiding from a --severity critical query. Lowering it back is a deliberate update, never an accident of the last report.

• A card that was fixed and comes back is a regression, not a duplicate. Re-filing reopens the same card and records the regression on it, so one defect's whole history stays on one row.

• A decision stays decided. Re-filing something already dismissed as wont_fix or not_a_bug does not quietly reopen the argument. It files a new card pointing back at the dismissal, so the recurrence is visible and the original ruling survives.

• The place in the code outlives the edit. When a report names a line range, the location is anchored at filing time from git. After the file is edited around it, anchor_resolve reports where that code went — moved, with the new line numbers — instead of pointing at whatever now occupies the old line.

• Parallel agents don't collide. claims_claim gives one agent a card and refuses the second, naming who holds it and from which repo, so two agents cannot silently fix the same thing. Closing a card releases the claim in the same transaction.

• Findings can be related and grouped. Cards link to each other, and a similarity report proposes families of near-duplicate findings as a dry run you inspect before merging anything.

• It tells you when it could not look. A tracker it cannot read, a file it cannot stat, an anchor whose card never named a code span — each of these comes back as undetermined, with the reason, rather than as a confident wrong answer. codebugs where will tell you which tracker a command is actually bound to and which channel decided that, because a binding you cannot see is a binding you cannot debug.

Underneath that, it is durable memory across sessions: findings survive the conversation that produced them, requirements are checked against the code that claims to implement them, blocked work resurfaces when its dependency resolves, and a release knows what is still stranded on a branch.

codebugs is one SQLite database (.codebugs/findings.db). Modules are self-registering, and the running server reports its own tool catalogue — the module table below is the set of them.

Install

# Global install (recommended)
pipx install codebugs

# Or with pip/uv
pip install codebugs

Setup

Create the tracker

Run this once per project, in the project root:

codebugs init

This creates .codebugs/findings.db. init is the only command that creates the .codebugs/ directory — every other command discovers an existing one by walking up from the current directory (unless you point it somewhere explicitly, see below), and refuses with an actionable error if there is none. That refusal is deliberate: silently creating an empty database is how findings go missing.

There is one deliberate exception, and it is worth stating precisely because it looks like the rule being broken. The upward walk treats an existing .codebugs/ directory as the opt-in, so if that directory is there but holds no findings.db, the next command creates the database inside it rather than refusing. The common way to end up in that state is an interrupted init — the directory is created before the database — and self-healing on the next command is more useful there than demanding a second init.

A tracker you name explicitly is held to the stricter rule. --repo, --tracker-root and $CODEBUGS_ROOT must resolve to a directory that actually contains findings.db; a .codebugs/ without one is refused, and the message names which channel pointed there. The difference is about evidence: standing inside a directory says something about where you are, while a named path is an assertion that can be mistyped, or exported into a shell days ago and inherited by an unrelated process. That is exactly where a silent empty tracker does the most damage.

Two consequences worth knowing:

• Run init at the project root, not in a subdirectory. Discovery binds to the nearest .codebugs/, so a nested tracker hides the project's real one from everything beneath it. init refuses to do this unless you pass --force.

• Git worktrees share the main repo's tracker. A worktree's .git is a file pointing at the main repo, which discovery follows — so findings filed from a worktree land in the project's database, not in a throwaway that dies with the worktree. init refuses to run inside a worktree for the same reason; run it in the main checkout.

  Two layouts are the exception: if the main repo is bare or was created with --separate-git-dir, git records no path back to a main checkout — its own git worktree list reports the git directory instead. Discovery usually refuses those with an explicit error rather than guessing. There is one case it cannot detect: a --separate-git-dir repo whose git directory is itself named .git looks exactly like a normal checkout, so discovery binds to the directory holding the git dir instead of the real checkout, silently. Nothing local can distinguish the two — git reports that directory as a valid work tree as well — so the remedy is to state the root explicitly (below). Run init in the main checkout before creating worktrees.

Pointing codebugs at a specific tracker

Discovery is a heuristic, so it has an override:

codebugs --tracker-root /path/to/project query   # this invocation only
export CODEBUGS_ROOT=/path/to/project            # this shell and anything it spawns

Resolution order, most specific first: a command's own explicit path argument (--repo, where a command has one) → --tracker-root → CODEBUGS_ROOT → walking up from the current directory. A per-command argument outranks a declaration because it names one operation's target, while a declaration is process-wide.

codebugs where     # show the current binding and which channel decided it

where is a diagnostic, not a precedence level: it prints the resolved root, the database path, and the channel — the fastest way to check that a command is about to read the tracker you think it is.

One thing is true of every declared root, for every command except init: a root that contains no .codebugs/ is a hard error, never a new tracker. The value may be a stale export inherited from another shell, and silently creating an empty database there is how findings go missing.

init is the exception, and it treats the two channels differently, because they carry different evidence. --tracker-root DIR init creates the tracker in DIR — the flag is typed on the command line being run, so it is an assertion about this invocation, exactly like a path argument. CODEBUGS_ROOT is ignored by init, which creates where you are standing, and warns that later commands will read somewhere else: an environment variable exported days ago and inherited by an unrelated process must never conjure a tracker in a directory you are not in. Whenever the tracker init created is not the one the next command would read, it says so on stderr, since otherwise it would report success for a dead end.

CODEBUGS_ROOT is inherited by every subprocess, so export it only when you mean "this shell works on that project". For one-off use across projects, prefer --tracker-root.

Claude Code (MCP)

Add to ~/.claude.json (global) or .mcp.json (per-project):

{
  "mcpServers": {
    "codebugs": {
      "command": "codebugs-mcp"
    }
  }
}

The database lives at .codebugs/findings.db, discovered by walking up from the server's working directory — each project gets its own. Run codebugs init in the project first (see above), or every tool call will fail with "no .codebugs/ found".

The server connects lazily, per tool call, so it starts successfully even when no tracker is reachable. At startup it writes a diagnostic to stderr — which MCP clients log — if discovery failed, or if a root was declared rather than discovered; on the ordinary path it says nothing. It never refuses to start: a project directory that appears later must still work.

To pin one server to one tracker instead of deriving it from the working directory:

{
  "mcpServers": {
    "codebugs": {
      "command": "codebugs-mcp",
      "args": ["--tracker-root", "/path/to/project"]
    }
  }
}

Only do this when you want that server bound to a single project — the default cwd-derived behavior is what lets one registration serve many. Add .codebugs/ to your .gitignore.

Running Modules Independently

Use --mode to load only the tools you need:

{
  "mcpServers": {
    "codebugs": {
      "command": "codebugs-mcp",
      "args": ["--mode", "findings"]
    }
  }
}

Any module name from the module table below is a valid mode, and all — the default — loads everything. The CLI takes the same flag: codebugs --mode findings summary.

One asymmetry is worth knowing before you rely on it: usage is a CLI-only mode. It registers a command but no MCP tools, so codebugs --mode usage usage works, while codebugs-mcp --mode usage does not start at all — the server does not accept that value, and refuses it with argument --mode: invalid choice: 'usage' and exit code 2. It is not a server that runs with an empty catalogue; there is no server. The two lists of accepted modes are compared with each other, and with the module table below, by the test suite — so a mode that appears on one surface and not the other turns a test red rather than surprising you here.

Other MCP Clients

Any MCP-compatible client can connect to codebugs-mcp via stdio transport.

The modules

Every name in this table is a valid --mode value.

Modules are self-registering — adding a new one is local to its own file. See docs/superpowers/specs/ for the architecture history.

Not every MCP tool has a CLI verb. The milestones module is where the gap is widest: milestone_create, milestone_update, milestone_add_item, milestone_move_item, milestone_set_status, milestone_defer, milestone_close, triage_dismiss, triage_promote, pull_next and release_item are reachable through MCP only. From a terminal you can inspect a release, but you cannot create one or pull work from it.

Quick tour

Findings — log it, never re-discover it

MCP tools:

staleness_check lives in the provenance module rather than here, and anchor_resolve in loc; both are listed in the module table. The tools behind the opening section — anchors, similarity, relations and grouping — have their own entry under Identity, location and grouping below.

CLI:

codebugs add -s high -c n_plus_one -f src/api.py -d "Query in loop at line 42" --new-category
codebugs summary
codebugs query --status open --severity critical
codebugs update CB-1 --status fixed --append-note "Fixed in PR #42"
codebugs categories

--new-category is needed the first time a category is used, and only then. A category the tracker has never seen is refused, naming the closest existing spellings, because the common way a category set fragments is a typo — n-plus-one beside n_plus_one — and a fragmented category set is what makes categories stop revealing patterns. Once n_plus_one exists, later findings use it without the flag. Spelling is normalized on the way in, so hyphens, spaces and case do not mint twins.

--append-note adds to a finding's notes; --notes replaces them wholesale. Prefer --append-note when recording investigation history — --notes will discard whatever was there, which is usually not what you want on a finding others have been working.

When a new finding is added, the milestones auto-router automatically attaches it to stream/triage (or stream/security when severity=critical and category starts with security:). The finding and its triage entry land in the same transaction.

Requirements — verify what shipped, surface contradictions

MCP tools:

CLI:

codebugs reqs-import REQUIREMENTS.md
codebugs reqs-summary
codebugs reqs-verify
codebugs reqs-query --status Implemented --priority Must
codebugs reqs-update FR-090 --status Superseded --notes "Replaced by vault architecture"
codebugs reqs-export REQUIREMENTS.md

Blockers — "X is blocked by Y", with auto-unblock

MCP tools:

Triggers come in three flavors: entity_resolved (waits for another finding/requirement to reach a terminal state), date (unblocks on a specific datetime), and manual (operator signal). When you mark a finding fixed, every blocker that was waiting on it auto-unblocks and surfaces in the next blockers_check.

Milestones — release containers + standing streams + capacity-aware pull

MCP tools:

Four seed milestones are created automatically:

• stream/triage — inbox for unsorted findings (default destination)

• stream/maintenance — deferred / boy-scout work

• stream/security — urgent fixes (preempts release work)

• release/1.1 — first post-1.0 release

pull_next priority order: stream/security > release/* (earliest target_date first) > stream/triage > stream/maintenance. Within a milestone: priority ASC, then created_at ASC.

Eligibility: item is open, no active blockers (skipped for item_kind='external'), acceptance required for size='large', and a large bug in a release milestone must declare linked_frs whose ids resolve to rows in requirements. Concurrent calls from multiple agents are atomic — claims are serialized via BEGIN IMMEDIATE.

CLI:

codebugs milestone-list
codebugs milestone-status release/1.1
codebugs triage-inbox
codebugs wip-status
codebugs milestone-audit --milestone release/1.1

A typical autonomous-agent loop:

# 1. Agent claims the next eligible item.
item = pull_next(agent_id="agent-A", capacity={"large": 1, "small": 2, "triage": 5})

# 2. (Optional) flag a feature branch.
mark_branch_only(item_ref=item["item_ref"], branch_name="feat/CB-1234")

# 3. After integration, mark it done with the commit SHA.
mark_integrated(item_ref=item["item_ref"], commit="abc123…")

# 4. Free the agent's capacity slot.
release_item(item_ref=item["item_ref"], status="done")

Closing a release runs the close-gate: unfinished, branch-only, and blocker-gated items refuse to let the milestone ship. force=True (with a logged reason) overrides — but stream/* milestones cannot be closed, even with force.

Sweeps — batch iteration with recurrence-aware lifecycles

MCP tools:

codebugs sweep-create --name lint-pass --batch-size 5
codebugs sweep-add lint-pass src/*.py --tags critical
codebugs sweep-next lint-pass
codebugs sweep-mark lint-pass src/api.py
codebugs sweep-status lint-pass

With a custom lifecycle (e.g. for retro findings):

codebugs sweep-create --name retro-findings \
    --lifecycle DETECTED,CONFIRMED,ESCALATED,RESOLVED,DROPPED \
    --terminal-states RESOLVED,DROPPED
codebugs sweep-add retro-findings finding-2026-04-todo-bypassed --tags silent_abandonment
codebugs sweep-mark retro-findings finding-2026-04-todo-bypassed --state CONFIRMED
codebugs sweep-archive-items retro-findings --state RESOLVED --older-than 30d

Bench — performance snapshots over time

MCP tools:

Merge — parallel-agent merge serialization

MCP tools:

Identity, location and grouping

The opening section's claims about identity, location and grouping rest on loc, similarity, relations and grouping, and none of them had an entry here before.

An anchor needs the filing to name a code span. The location is captured from meta.line / meta.lines (and a few equivalent spellings) at the moment the finding is filed. A report that names only a file has nothing to anchor, and anchor_resolve says so — unknown, reason no_grammar — instead of guessing a line. Most findings in practice name no span, so this is the ordinary case, not an error:

$ codebugs anchor-resolve --finding-id CB-2 --repo . --json
      ...
      "anchor": {
        "status": "moved",
        "path": "src_api.py",
        "line": 25,
        "end": 27,
        "channel": "git",
        "reason": null,
        "survived": "3/3",
        ...

That card was filed against lines 5–7. Twenty lines were then inserted above it. The stored line numbers are stale; the anchor is not.

The ... are elision marks rather than output: this is one entry lifted out of a larger JSON document, and the anchor object carries one further key below survived — resolved_against, the root, commit and on-disk file the answer was computed against. Nothing is ever omitted because it was absent. Every key is present on every answer, so "reason": null here means there is nothing to explain, because there is an answer — never this field is missing.

How It Works

The Problem

AI code review sessions produce findings that get lost. Multiple agents working in parallel double-claim work. Requirements files drift. Releases lose track of what's in them.

The Solution

codebugs stores everything in one local SQLite database. AI assistants write findings, requirements, and milestone items as they discover them, then query the database in future sessions for instant context recovery. Concurrent agents coordinate via the same database — no race conditions, atomic claims.

Token savings: a summary call returns one small structured overview — counts by severity, the top categories, the hottest files — instead of the file reading and conversation replay it would otherwise take to re-establish the same picture.

Typical Workflows

Code review loop. This is the same loop the MCP server tells every client that connects, so the two cannot drift apart:

1. File the observation with add (or batch_add for several at once). Call categories first when you are unsure of the naming.

2. Read what came back before doing anything else. dedup_action says whether this created a new card, bumped or reopened an existing one, or refiled one already dismissed. attention is the server's own flag when your observation raised the card's severity or diverged from its stored category — a card you thought you were filing fresh but which the tracker already knew about, differently.

3. The code location is anchored at file time when the observation names one; anchor_resolve reports whether that anchor still points at live code.

4. Close the card with update CB-N --status fixed once it is actually fixed.

Working alongside other agents on the same tracker? Claim the card with claims_claim before starting and release it with claims_release when done, or two agents can end up fixing the same thing. Closing the card releases the claim on its own.

Each add also auto-routes the finding to stream/triage, and over time categories reveals systemic issues — "12 tz_naive_datetime fixed across 9 files → time for a lint rule."

Requirements (reqs_add, reqs_query, ...) are a separate, authored entity next to findings, and they have no deduplication. Do not file a requirement through add, or a defect through reqs_add.

Release loop:

1. Triage: AI calls triage_inbox → triage_dismiss non-bugs, triage_promote real items to release/1.1 (with linked_frs for the ones that need an FR row).

2. Execution: Each parallel agent calls pull_next(agent_id=..., capacity=...) → claims the next eligible item.

3. After landing: mark_integrated(item, commit) → release_item(item, status='done').

4. Close: milestone_close("release/1.1"). Refuses if anything is stranded on a branch; lists the offenders with the branch name.

Schema (highlights)

All tables share .codebugs/findings.db with flexible JSON columns. Schemas are additive — every module owns its tables, declares dependencies, and migrates additively.

Findings

Requirements

Milestones

Item kinds are bug (validated against findings), requirement (validated against requirements), or external (free-form, blockers skipped). The (milestone_id, item_kind, item_ref) unique constraint prevents double-attach.

Blockers

Sweeps

Killer features

Pattern detection over time

$ codebugs categories
category                  total  open  fixed
------------------------  -----  ----  -----
tz_naive_datetime         15     3     12
n_plus_one                8      2     6
missing_input_validation  6      4     2

If you keep fixing the same category → time for a lint rule. codebugs turns reactive bug-fixing into proactive prevention.

This is the view the category gate above protects. Had half those findings been filed as tz_naive_dt, the table would show two rows of 8 and 7 instead of one row of 15, and there would be no pattern to see. Normalization handles the punctuation-and-case twins on its own; the gate is what catches a genuinely different name for the same thing.

Requirements verification

reqs_verify catches documentation rot before it ships:

$ codebugs reqs-verify
Verified 3 requirements.

4 issue(s) found:

check   sev     id      message
------  ------  ------  -----------------------------------------------------------
ids     medium  --      Numbering gaps (5+): FR-007..FR-089, FR-091..FR-349
status  medium  FR-006  Must-priority requirement implemented without test coverage
status  high    FR-090  Description mentions 'superseded' but status is 'planned'
status  medium  FR-350  Must-priority requirement implemented without test coverage

Semantic requirements search

Store embeddings (caller generates vectors via any embedding API) and find related requirements semantically:

reqs_embed(req_id="FR-001", embedding=[0.1, 0.2, ...])
reqs_search_similar(query_embedding=[...], limit=5, min_similarity=0.3)

Float32 BLOB storage in SQLite; brute-force cosine similarity — fast for thousands of requirements.

Close-gate enforcement

milestone_close won't let you ship a release with work stranded on a branch. First, what the release looks like:

$ codebugs milestone-status release/1.1
release/1.1  (release, state=open)
  target: 2026-09-15 (19 days)
  First post-1.0 feature release. Target date set later.

Items: 3 total (3 open/in_progress, 0 done)

  By status:
    open              3
  By size:
    small             3

  Branch-only: CB-1
  Blocked: CB-2

Then closing it. milestone_close is one of the milestone tools with no CLI verb — this is the error the MCP tool returns, raised as a ValueError from the domain function, on a single line:

cannot close release/1.1: unfinished items (3): CB-1, CB-2, CB-3; branch-only items (1): CB-1@feat/CB-1; items with active blockers (1): CB-2  (use force=True with reason to override)

force=True with a logged reason overrides that. Streams do not have an override — milestone_close(id="stream/triage", force=True, reason="x") still refuses with streams cannot be closed (milestone=stream/triage), because they are permanent buckets rather than things that ship.

Requirements

• Python 3.11+

• One runtime dependency: mcp>=2.0.0,<3, for the server. The 2.0 floor is not cosmetic — server.py uses MCPServer, the class that replaced FastMCP in the 2.0 SDK, so an older mcp will not start.

• SQLite (bundled with Python)

Development

# Run tests
uv run --extra dev python -m pytest tests/ -q

# Lint — this is the gate
uv run --extra dev ruff check src/ tests/

pytest and ruff live in the dev extra, which uv run does not install by default, so --extra dev is not optional in a fresh clone.

ruff format is deliberately not run over this tree. Much of the existing code does not conform to it, so ruff format src/ tests/ would rewrite most of the repository in one commit. ruff check is the gate; formatting is left alone on purpose.

See CLAUDE.md for architectural rules and conventions.

License

MIT