# Reported issues for Madar

Pod holds 20 of 75 problems reported by people outside the maintainer team. Issues filed by the project's own owners, members and collaborators are excluded entirely — a maintainer's release checklist is not a warning to a prospective user.

Back to [Madar](/mcp/madar).

## Most discussed

### [P0] Complete artifact v2 cutover and tombstone the legacy graph.json contract

## Problem

PR B1 (`#657`) introduces and activates the canonical artifact v2 for new binaries, while retaining a fresh v1 compatibility mirror at `out/graph.json`. New readers prefer `out/graph.madar`; the legacy path is still a real v1 artifact.

The final cutover — demoting `graph.json` to a non-JSON tombstone so old binaries fail closed — is separated because it changes a large public and test contract that has nothing to do with graph storage. Bundling it into the storage PR would have made

[Read the thread](https://github.com/mohanagy/madar/issues/705) · 2026-08-15 · closed · 28 comments

### [P0][Architecture]: Build answer-ready semantic execution retrieval without source fallback

## Outcome

Replace declaration-only evidence retrieval with an answer-ready semantic execution index for statically provable JavaScript/TypeScript questions.

The caller-facing surface remains one MCP tool: `retrieve({ question, budget? })`. Madar must either return a complete ordered, claim-to-proof dossier or an exact non-ready result. Claude must not need repository Read/Grep/Glob/Bash fallback for supported workflow questions.

## Protected base and target

- Architecture baseline: protecte

[Read the thread](https://github.com/mohanagy/madar/issues/629) · 2026-07-30 · closed · 15 comments

### [P0][Index]: Add authenticated execution facts, ordering, persistence, and shared async channels

Parent: #629

## Outcome

Extend the canonical JavaScript/TypeScript index with compact authenticated execution facts and exact shared channel topology, without turning every AST expression into a graph node.

## Scope

- Add ordered body facts owned by callable/constant symbols: call, literal, condition, loop, parallel, return, throw, mutation and persistence.
- Store numeric AST order, bounded statement/evidence ranges, excerpt SHA-256, control frames, confidence and provenance.
- Retain only 

[Read the thread](https://github.com/mohanagy/madar/issues/632) · 2026-07-30 · closed · 12 comments

### [Core Reset]: Rebuild Thin Delivery as a lazy CLI, one-tool MCP, and verified Claude/Codex transport

## Status

**Ready for owner review. Not accepted, not activated, and no implementation is authorized by opening this issue.**

- Governing RFC: #577
- Protected planning base: `8efe41fc665fcea7e625dda0864a72ecf27a111b`
- Manifest handle: `thin-delivery`
- Completed prerequisite: #596, including absorbed #599
- Absorbed activation blocker: #567
- Required next step: explicit owner approval of this exact contract, followed by a governance-only activation PR

## User outcome

A JavaScript/TypeScri

[Read the thread](https://github.com/mohanagy/madar/issues/602) · 2026-07-25 · closed · 9 comments

### P0: Codex installer writes project MCP config that Codex CLI never loads

## Problem

`madar codex install` currently writes the managed MCP block to `<repo>/.codex/config.toml`. A real Codex CLI activation test against a fresh OpenStatus worktree showed that Codex CLI 0.144.5 does not load that project-local file:

```text
cd <repo>
codex mcp list
# lists only ~/.codex/config.toml servers; no madar server
```

The installed `AGENTS.md` guidance is visible, so Codex attempts the local `madar pack` fallback instead of the strict MCP tool. The project config is therefor

[Read the thread](https://github.com/mohanagy/madar/issues/567) · 2026-07-17 · closed · 6 comments

### [Bug]: beta.2 full-flow retrieval remains prompt-sensitive and misses async pipeline handoffs

## Summary

`@lubab/madar@0.40.0-beta.2` fixed the earlier hard failure where Claude treated the graph as unavailable, but a real natural-language report-flow prompt remains unstable. A shorter phrasing returns useful pipeline anchors and edges; a semantically equivalent, more explicit phrasing over-selects downstream assembly details, omits the orchestration/research/DB-sync skeleton, and forces Claude into broad filesystem inspection.

This is one focused ranking/recovery regression. It is not

[Read the thread](https://github.com/mohanagy/madar/issues/622) · 2026-07-29 · closed · 5 comments

### Create a public distribution and launch checklist for adoption channels

## Problem
Madar is free and open source, but free does not create distribution. Developers still need to encounter it in the right context, with enough proof and a low-risk first step.

Current adoption risk:
- npm/GitHub discovery alone is weak.
- Search results and external listings may lag behind the Madar rename.
- Community posts need concrete receipts, not abstract architecture.

## Scope
Create a repeatable launch/distribution checklist for each meaningful release or benchmark milestone.

[Read the thread](https://github.com/mohanagy/madar/issues/474) · 2026-06-02 · closed · 5 comments

### [P0] Prevent agent turn explosion after Madar context injection

## Priority
P0

## Type
Bug / Agent efficiency

## Area
Agent integrations, MCP, context packs, compare, prompt guidance

## Problem
Madar can produce useful context but still fail the product promise if the agent treats the pack as extra context and continues broad repo exploration.

Recent benchmark evidence showed both outcomes:

- Good run: FounderCommandCenter auth prompt with non-SPI graph: baseline 19 turns → Madar 4 turns, with lower provider-reported input tokens and lower cost.
- Bad r

[Read the thread](https://github.com/mohanagy/madar/issues/314) · 2026-05-25 · closed · 4 comments

## Most recent

### [P1] Reduce artifact v2 load latency without weakening identity verification

## Problem

PR B1 (#657) accepted a reviewed exception to the 2× load target:

| | Value |
|---|---:|
| Base (v1 loader) median | 308.19 ms |
| Candidate (v2 loader) median | 675.69 ms |
| Ratio | 2.192× |
| Artifact size ratio | 1.894× |
| Residual per-byte overhead | ~1.157× |

Receipts: `docs/benchmarks/2026-08-15-graph-b1-artifact-v2/` — initial gate, post-remediation gate, and the maintainer decision.

The residual is not a redundant loop. Four profiling rounds removed duplicate receipt acc

[Read the thread](https://github.com/mohanagy/madar/issues/706) · 2026-08-15 · open · 0 comments

### [P1] Publish npm releases from exact protected commits with provenance and packaged host verification

Parent: #652
Depends on: #673, #674, #675

## Outcome

Remove the maintainer workstation as a release trust boundary.

Publish Madar to npm only from an exact protected release commit through an auditable workflow that produces provenance, package identity, SBOM/checksums where supported, Registry verification, clean-install execution, artifact migration tests, and packaged host smoke receipts.

## Problem

The existing release process has strong CI and packaging checks, but npm publication rema

[Read the thread](https://github.com/mohanagy/madar/issues/676) · 2026-08-11 · open · 0 comments

### [P1] Establish a versioned compatibility, support, and deprecation policy

Parent: #652
Depends on: #674

## Outcome

Define exactly which Madar interfaces and artifacts are supported, how they are versioned, what backward compatibility means, how breaking changes are introduced, and how users migrate safely.

The policy must cover Pack, graph artifacts, CLI, MCP, installers, supported runtime/toolchain versions, and the unresolved question of whether Madar is an application-only CLI/MCP product or a supported importable TypeScript library.

## Problem

Madar currently

[Read the thread](https://github.com/mohanagy/madar/issues/675) · 2026-08-11 · open · 0 comments

### [P1] Rationalize supported, advanced, and maintainer CLI/MCP surfaces

Parent: #652
Depends on: #666, #673

## Outcome

Define a small, coherent supported product surface for ordinary Madar users while preserving advanced and maintainer capabilities behind explicit tiers and documented compatibility windows.

Reduce command/tool-selection ambiguity, make help/onboarding/doctor behavior consistent, and ensure the normal path centers on trusted context compilation rather than a large taxonomy of graph and maintenance operations.

## Problem

Madar currently combines:

[Read the thread](https://github.com/mohanagy/madar/issues/674) · 2026-08-11 · open · 0 comments

### [P1] Establish a framework-adapter qualification and conformance contract

Parent: #651
Depends on: #662, #668

## Outcome

Define a stable internal contract by which language/framework adapters declare capabilities, emit semantic facts and evidence, report uncertainty, and prove qualification against independent positive and negative fixtures.

The contract must make support claims testable and prevent a new adapter from appearing complete while silently omitting dynamic, generated, unsupported, or unresolved behavior.

This is an internal extension contract. It is no

[Read the thread](https://github.com/mohanagy/madar/issues/669) · 2026-08-11 · open · 0 comments

### [P1] Qualify and stage a context-pack-first default for supported agent hosts

Parent: #650
Depends on: #663, #664, #661

## Outcome

Make Madar’s trustworthy task-to-context workflow naturally discoverable and selected by supported coding-agent hosts.

Stage context-pack-first defaults only after the pack, integrity, uncertainty, and answerability contracts are qualified. Preserve explicit compatibility/advanced profiles and a rollback switch.

## Confirmed baseline mismatch

At v0.32.1:

- `core` exposes graph/retrieval tools but not `context_pack`;
- `strict` exposes `c

[Read the thread](https://github.com/mohanagy/madar/issues/666) · 2026-08-11 · open · 0 comments

### [P1] Add an end-to-end evidence, ranking, omission, and compaction trace

Parent: #650
Depends on: #664

## Outcome

Make a compiled context pack explainable from task interpretation through final answerability.

Maintainers must be able to determine why a file, symbol, fact, snippet, claim, execution step, caveat, or verification target was included, excluded, downgraded, compacted, or omitted—and which policy/budget/integrity decision caused it.

## Problem

Madar already exposes useful diagnostics, but there is no single trace connecting:

```text
task interpretati

[Read the thread](https://github.com/mohanagy/madar/issues/665) · 2026-08-11 · open · 0 comments

### [P1] Introduce one canonical internal context-compilation seam

Parent: #650
Depends on: #663

## Outcome

Establish one canonical internal operation that converts a task, repository evidence, budget, and policy into an audited Pack v1-compatible result.

CLI, MCP, prompts, and command wrappers must delegate to this operation rather than independently reconstructing retrieval, compaction, evidence, answerability, or rendering contracts.

This issue is a controlled façade-and-migration effort, not a clean-architecture rewrite.

## Problem

Current context beh

[Read the thread](https://github.com/mohanagy/madar/issues/664) · 2026-08-11 · open · 0 comments

### [Epic] Phase 3 — Simplify the supported product surface and protect releases

Parent: #648
Depends on: #651

## Outcome

Make Madar straightforward to install, discover, operate, debug, upgrade, and trust as a released product.

The supported product surface should center on one trustworthy task-to-context workflow plus bounded expansion/status operations. Advanced graph exploration, exporters, benchmark controls, and maintainer operations may remain available, but they must not obscure the normal user path or silently expand the compatibility contract.

## Problem

Madar

[Read the thread](https://github.com/mohanagy/madar/issues/652) · 2026-08-11 · open · 0 comments

### [Feature]: rebuild fails permanently on ENOENT when git changes branches, blocking all MCP queries

### Problem to solve

When `madar watch` is running and a git operation removes files (branch switch,
  merge, rebase, reset), the incremental rebuild throws `ENOENT` and the watcher
  latches into `status: "failed"` permanently. From that point `serve --auto-refresh`
  rejects **every** query, even though a valid `out/graph.json` is still on disk.

  A file disappearing mid-rebuild is a normal race with git — not an error condition.

  **Version:** `@lubab/madar@0.32.0` (global install, Node 24

[Read the thread](https://github.com/mohanagy/madar/issues/645) · 2026-08-11 · closed · external user · 2 comments

### [Bug]: doctor reports attention required for a healthy Claude-only setup

## Summary

In `0.40.0-beta.1`, `madar doctor` reports `attention required` even when the graph is ready and the requested Claude workspace registration is exact. The only reported problem is that an optional Codex registration is absent.

This makes a healthy Claude-only installation look broken.

## madar version

`0.40.0-beta.1` from an npm global install.

## Command or workflow

```text
madar generate .
madar install claude
madar doctor
```

Observed diagnostic shape:

```text
[madar doctor

[Read the thread](https://github.com/mohanagy/madar/issues/619) · 2026-07-28 · closed · 1 comment

### [Bug]: 0.40 beta misses a known JS/TS flow that 0.32 retrieves

## Summary

`@lubab/madar@0.40.0-beta.1` completes indexing and is callable through an exact workspace-local Claude registration, but a natural flow question returns `missing` after three retrieval attempts and forces a filesystem fallback. The current official `0.32.0` produces a materially better, nearly complete answer for the same repository and task.

This is a retrieval regression, not an installation or unsupported-language problem.

## madar version

`0.40.0-beta.1` from an npm global in

[Read the thread](https://github.com/mohanagy/madar/issues/618) · 2026-07-28 · closed · 2 comments

The remaining reports are on [the project's issue tracker](https://github.com/mohanagy/madar/issues).