Honestcode

The problem

AI coding agents hallucinate. They invent function names, fabricate library APIs, and call methods that do not exist.

HonestCode is the layer that checks what the agent just wrote against the repository it is supposed to be grounded in.

Agent  ↓writes code  ↓HonestCode  ↓evidence  ↓Agent fixes  ↓verified code

No LLM calls. No network. Pure AST + symbol resolution.

Demo

The agent was asked to "add authentication using the existing UserClient." It generated:

client.refresh_token()

The real project only has:

class UserClient:    def refresh(self): ...    def refresh_access_token(self): ...

HonestCode catches it deterministically:

✗ verification failed — 1 issue in app/login_broken.py
  [invented_api] app/login_broken.py:9  `UserClient.refresh_token()` does not exist.  did you mean: refresh_access_token()?  available methods:    - refresh()    - refresh_access_token()  confidence: deterministic · action: revise

Run it yourself:

cd demos/invented-apipython run.py

Why not just use Ruff / Pyright?

HonestCode is not a replacement — it is a verification layer that answers a specific question: did the agent's code actually come from this repository?

The key difference is repository grounding. Ruff checks the file. Pyright types the call. HonestCode checks whether the call resolves to a real symbol that exists in the codebase the agent is editing. (HonestCode does not do full type inference — its structural checks catch mismatched arities, iterating None, and calling constants, not general type errors.)

What a call site is checked against depends on how the receiver's type is known: an annotation (client: UserClient), a constructor assignment (client = UserClient()), self/cls, and an inline constructor (UserClient().refresh()) are all resolved. A receiver produced by a factory function (make_client().refresh()) is not judged — function signatures carry no return type in the surface model, and guessing there would trade a missed finding for a false one.

Install

One line — no clone, no virtualenv:

claude mcp add honestcode -- uvx honestcode-mcp

uvx fetches the server into an isolated environment on first use; install uv first if you do not have it. From a checkout (to hack on the code, or if you prefer not to use uv):

git clone https://github.com/Fengrru/honestcodecd honestcodepip install -e .claude mcp add honestcode -- honestcode-mcp

Cursor / VS Code / Claude Desktop: the honestcode install command probes those clients' config files and registers the server for you:

honestcode install            # all detected clientshonestcode install --client cursor --dry-run   # preview without writing

Requires Python >= 3.10. 100% local.

30-second setup with Claude Code

# Add HonestCode as an MCP serverclaude mcp add honestcode -- honestcode-mcp

Or add it manually to your Claude Code config (~/.claude.json on macOS/Linux, %USERPROFILE%\.claude.json on Windows):

{  "mcpServers": {    "honestcode": {      "command": "honestcode-mcp",      "args": []    }  }}

Restart Claude Code. You now have two verification tools available:

• scan_file(path) — legacy tool, returns issues plus the new evidence shape.
• verify_file(path) — agent-facing tool, returns status, findings with evidence, and a human-readable text summary.

Try prompting Claude:

Add a login endpoint using the existing UserClient, then run verify_file on thefile you just wrote and fix anything it reports.

HonestCode auto-discovers the project root from the file path, indexes the codebase, and returns grounded evidence.

Making it automatic (no prompt discipline required)

MCP tools only run when the model chooses to call them. To make verification unconditional, wire the hook in scripts/claude_hook.py into Claude Code's PostToolUse event — every .py file the agent writes is verified, and a finding blocks the edit with the evidence fed straight back into the conversation:

{  "hooks": {    "PostToolUse": [      {        "matcher": "Write|Edit",        "hooks": [          {            "type": "command",            "command": "python /absolute/path/to/honestcode/scripts/claude_hook.py"          }        ]      }    ]  }}

Put that in ~/.claude/settings.json (all projects) or .claude/settings.json (one project). The hook never blocks on its own failures: unreadable input, non-Python files, and verification crashes all exit silently, so it can never turn into a wall that blocks every edit.

Real example

# auth/client.pyclass UserClient:    def refresh_access_token(self): ...
# auth/login.pyfrom auth.client import UserClient
def login(client: UserClient):    client.refresh_token()  # invented API

>>> from honestcode.mcp.tools import verify_file>>> verify_file("auth/login.py"){  "status": "fail",  "file": "auth/login.py",  "findings": [    {      "line": 7,      "kind": "invented_api",      "symbol": "refresh_token",      "owner": "UserClient",      "message": "`UserClient.refresh_token()` does not exist.",      "evidence": {        "available_methods": ["refresh", "refresh_access_token"],        "did_you_mean": "refresh_access_token"      },      "confidence": "deterministic",      "action": "revise"    }  ],  "text": "✗ verification failed — 1 issue in auth/login.py\n  ..."}

Agent output protocol

verify_file returns evidence an agent can act on directly, not just an error message:

{  "status": "fail",  "file": "auth/client.py",  "line": 42,  "kind": "invented_api",  "symbol": "refresh_token",  "owner": "UserClient",  "message": "UserClient.refresh_token() does not exist.",  "evidence": {    "available_methods": ["refresh", "refresh_access_token"]  },  "confidence": "deterministic",  "action": "revise"}

How it works

1. Auto-index the project (or reuse the cached symbol index).
2. Parse the target file with the standard-library ast module.
3. Resolve every call site to a concrete symbol:
   • infer the receiver type from annotations, constructors, and imports;
   • reconstruct the class's member surface from the repository AST;
   • mark the surface as unknown when a base class cannot be resolved or the class defines __getattr__.
4. Emit findings with kind, owner, evidence, confidence, and action.

The loop is deterministic and auditable.

Scope and limitations

The verifier stays silent wherever it cannot prove absence — that is the feature, but it means detection has known boundaries:

• Decorated classes are not judged for invented members. Any class decorator (including @dataclass) may add or rewrite members, so invented method calls on such classes are not reported. Undecorated classes are fully checked.
• Standard-library member calls are not checked. from typing import TypeVarr (an invented import) is caught, but os.path.joinn() (an invented call on a real module) is not — third-party and project members are the covered surface.
• Factory-function receivers are not judged. make_client().refresh() is skipped; annotations, constructor assignments, self/cls, and inline constructors (UserClient().refresh()) are resolved.
• undefined_symbol carries confidence: high, not deterministic. Only member-surface claims (invented APIs on a fully enumerated class) are provable from the repository; name absence rests on the syntactic binding model and keeps the lower confidence. Variables reassigned across types in one scope are treated as unknown and never judged.
• Graph tools are anchored to the last indexed project. scan_file / verify_file resolve the project root per file (including monorepo sub-projects); find_dead_code, explore_call_graph, search_code and friends operate on the project most recently passed to index_project.
• check_api imports the base package on demand — but only when it is genuinely installed (or stdlib). A bare file dropped into the workspace is never imported, so agent-authored code cannot be executed server-side through this tool.

MCP tools

By default scan_file and verify_file are exposed. Set HONESTCODE_TOOLS=all to enable the full set, including:

Dependency manifests (requirements.txt, pyproject.toml) are declared automatically on the first scan, so imports from a dependency that is not installed in the MCP server's interpreter are trusted rather than reported as invented. Set HONESTCODE_AUTO_DEPS=0 to disable that.

See the old tool reference below for the complete list.

Crash-safe side effects

verify_file checks code before it runs. guards/ covers what happens after: an agent that retries a tool call can charge a card twice, and no durable execution engine prevents it — Temporal and Restate both guarantee deterministic replay of orchestration and leave exactly-once external effects to the application.

from honestcode.guards import EffectStore, SideEffectGuard, durable_scope
guard = SideEffectGuard(session_id="run-42", store_path=".agent/effects.db")
outcome = guard.run(    "charge",    target="cus_123",    payload={"cents": 4200},    execute=lambda: charge_card("cus_123", 4200),    approve_irreversible=True,    scope=durable_scope("order-4711", "charge", "cus_123"),)
outcome.executed     # False on the retry — a receipt, not a second chargeoutcome.effect_id    # stable across a restart, because the key is on disk

The interesting part is what scope does. Pass it and the effect's identity comes from the scope rather than the session — which is the difference between a retry being recognised and a customer being charged twice, because a restart is a new session id. Pass nothing and the guard stays session-scoped.

Refusing is the default direction: an unrecognised action is treated as irreversible rather than idempotent, and irreversible effects need explicit approval. A run that dies between the call and the record leaves the effect executing, and the next attempt reports the outcome as unknown instead of guessing — which is what stops the retry. Scoped effects are recovered with guard.recover(scope=durable_scope(...)), and guard.effects(scope=...) / guard.summary(scope=...) report them; the records live under a namespace derived from the scope so a restart can still find them.

For a framework, integrations/ binds the guard to pydantic-ai, where a committed effect replaces the tool result instead of re-running the tool. See honestcode/integrations/README.md.

Benchmark

HonestCode includes four benchmark suites that run automatically in CI on every push and pull request to main:

Agent-accuracy benchmark

benchmarks/agent_accuracy/ is a deterministic, LLM-free benchmark that measures how well HonestCode catches common agent hallucinations (recall).

Each task is a tiny agent episode: the agent writes a broken file, HonestCode verifies it, then the file is replaced with the fix and verified again.

cd benchmarks/agent_accuracypython run.py

Current results (6 tasks, deterministic verification):

Summary: precision 1.0, recall 1.0, F1 1.0, false-positive rate 0.0, median verify time 4–7 ms on a laptop.

The per-task milliseconds above are one run's output, not a guarantee — they move a few tens of percent between runs on the same machine. What is stable, and what the benchmark is gating on, is the first four columns plus the summary's counts.

The summary line is python benchmarks/agent_accuracy/run.py verbatim rather than a figure copied into this file, because a timing quoted here goes stale the moment the code gets faster — as one did. Reproduce it with --format markdown.

The cross_file_new_symbol task pins the agent-session flow (write a module, then a file importing it — the index must be fresh, never stale); the dep_alias_and_pypi_name task pins dependency grounding through aliased submodules and PyPI-named requirements (PyYAML → yaml).

Real-world false-positive benchmark

benchmarks/realworld_fp/ measures precision on real code: it scans a vendored snapshot of psf/requests (19 files, ~6.4k LOC, Apache-2.0) and asserts HonestCode reports zero findings. Real code exercises the constructs synthetic tasks never do — stdlib imports, keyword-call style, @staticmethod members, classes inheriting stdlib ABCs, compat re-export modules — and any finding on this corpus is, by construction, a false positive.

The same run also scans benchmarks/realworld_fp/regressions/: one tiny project per false-positive class found in the wild (a receiver rebound across types, with ... as bindings, monorepo sub-project imports, local binding forms). Each scenario fails the benchmark by name if the regression returns.

python benchmarks/realworld_fp/run.py# files: 25  findings: 0  elapsed: 0.7s (~26 ms/file)

The finding count is the claim; the elapsed time is indicative, and varies with the machine.

Value-drift benchmark

benchmarks/value_drift/ asks a different question: when a formatting-tolerant edit matcher resolves a search string, did the edit land on the region the caller named, or silently on a near-miss sibling that carries a different value? It runs ten scenarios across a guarded matcher cascade and an unguarded one, and classifies each edit as correct, silent_drift, no_match, or wrong_span. The guarded matcher must produce the intended edit on every scenario; a no_match is a refusal, not a pass.

python benchmarks/value_drift/run.py            # human-readablepython benchmarks/value_drift/run.py --json     # machine-readable (status on stderr)

Performance benchmark

scripts/benchmark.py measures the latency of core operations (index, scan, graph, dead code detection, similarity search) on the repository itself.

python scripts/benchmark.py                          # text outputpython scripts/benchmark.py --format markdown        # table for READMEpython scripts/benchmark.py --repo psf/requests      # benchmark a real-world repo

All four benchmarks run in CI (.github/workflows/ci.yml) as separate jobs: benchmark-accuracy, benchmark-realworld, benchmark-value-drift, and benchmark-performance. The accuracy job gates on precision and recall staying at 1.0, the real-world job on any finding or failed scan, and the value-drift job on every guarded scenario landing correctly. The performance job reports numbers rather than gating on them.

See benchmarks/agent_accuracy/README.md for the dataset format and how to add new accuracy tasks.

Architecture

honestcode/├── verify/        # Evidence protocol + repository-grounded verification├── mcp/           # MCP server layer├── honest/        # Symbol index + project binding├── graph/         # Persistent call graph (SQLite)├── structure/     # AST extraction├── sandbox/       # Sandboxed execution├── guards/        # Crash-safe side effects (stdlib-only, extractable)├── integrations/  # Framework bindings — pydantic-ai today└── cli.py         # Command-line interface

verify_file is the agent interface. scan_file is exposed alongside it by default and returns both the new evidence shape and the legacy issues list.

Two of those directories answer a different question from the rest. Everything above guards/ asks is this code consistent with the repository. guards/ asks did this effect already happen — an agent that retries a tool call after a crash can charge a card twice, and durable-execution engines leave exactly-once external effects to the application. integrations/ binds the guard to a framework; pydantic-ai's wrap_tool_execute is the first, and a tool result is replaced rather than re-run when the effect has already been committed.

guards/ is deliberately stdlib-only and does not depend on this package's protocols, so it can be extracted into its own distribution. tests/test_guards.py enforces that with an AST import check.

Roadmap

v0.1 — Repository grounding (done)

• symbol / API / function / class / method / import / call verification
• invented API detection with structured evidence
• auto-index on scan_file

Beyond the roadmap — crash-safe side effects (now)

Not a verification layer, so it has no place in the v0.2/v0.3 sequence below: guards/ makes an effect idempotent across a crash, and integrations/ binds it to pydantic-ai. It answers "did this already happen", which is a runtime question rather than a static one.

v0.2 — Semantic contract verification

• function expects UserID, agent passes User → suspicious

v0.3 — Execution verification

• static verification → tests → runtime evidence

v1.0 — Verification Runtime for Coding Agents

                 Coding Agent                      │              ┌───────▼───────┐              │   HonestCode  │              │ Verification  │              │    Runtime    │              └───────┬───────┘                      │       ┌──────────────┼──────────────┐       ↓              ↓              ↓   Static          Semantic       Runtime Verification     Verification   Verification       │              │              │       └──────────────┼──────────────┘                      ↓                   Evidence                      ↓                     Agent

Development

git clone https://github.com/Fengrru/honestcode.gitcd honestcodepython -m venv .venv.venv\Scripts\activate  # Windows# source .venv/bin/activate  # macOS/Linuxpip install -e ".[dev]"pytestruff check honestcode tests scripts

Run benchmarks locally:

# Accuracy benchmarkpython benchmarks/agent_accuracy/run.py
# Performance benchmarkpython scripts/benchmark.py

See CONTRIBUTING.md for pull request guidelines.

Security

See SECURITY.md for vulnerability reporting.

Telemetry

HonestCode collects no telemetry. There are no analytics libraries, no background services, and no phone-home endpoints.

License

MIT - Copyright (c) 2026 HonestCode Team

MCP Registry: mcp-name: io.github.Fengrru/honestcode