AGENTS.md — franken_whisper

Guidelines for AI coding agents working in this Rust codebase.

RULE 0 - THE FUNDAMENTAL OVERRIDE PREROGATIVE

If I tell you to do something, even if it goes against what follows below, YOU MUST LISTEN TO ME. I AM IN CHARGE, NOT YOU.

RULE NUMBER 1: NO FILE DELETION

YOU ARE NEVER ALLOWED TO DELETE A FILE WITHOUT EXPRESS PERMISSION. Even a new file that you yourself created, such as a test code file. You have a horrible track record of deleting critically important files or otherwise throwing away tons of expensive work. As a result, you have permanently lost any and all rights to determine that a file or folder should be deleted.

YOU MUST ALWAYS ASK AND RECEIVE CLEAR, WRITTEN PERMISSION BEFORE EVER DELETING A FILE OR FOLDER OF ANY KIND.

Irreversible Git & Filesystem Actions — DO NOT EVER BREAK GLASS

Absolutely forbidden commands: git reset --hard, git clean -fd, rm -rf, or any command that can delete or overwrite code/data must never be run unless the user explicitly provides the exact command and states, in the same message, that they understand and want the irreversible consequences.
No guessing: If there is any uncertainty about what a command might delete or overwrite, stop immediately and ask the user for specific approval. "I think it's safe" is never acceptable.
Safer alternatives first: When cleanup or rollbacks are needed, request permission to use non-destructive options (git status, git diff, git stash, copying to backups) before ever considering a destructive command.
Mandatory explicit plan: Even after explicit user authorization, restate the command verbatim, list exactly what will be affected, and wait for a confirmation that your understanding is correct. Only then may you execute it—if anything remains ambiguous, refuse and escalate.
Document the confirmation: When running any approved destructive command, record (in the session notes / final response) the exact user text that authorized it, the command actually run, and the execution time. If that record is absent, the operation did not happen.

Branch Policy

Primary branch is main.
Do not reference master in docs/scripts.
If release instructions require sync, push main:master after main.

Project Mission

franken_whisper is a native Rust, memory-safe, high-performance speech system that merges ideas from:

legacy_whispercpp (fast local inference + streaming + VAD)
legacy_insanely_fast_whisper (GPU-first batching + ergonomic CLI)
legacy_whisper_diarization (forced alignment + diarization + punctuation restoration)

And integrates deeply with:

/dp/asupersync for cancel-correct orchestration
/dp/frankentorch and optionally /dp/frankenjax for accelerated compute backends
/dp/frankensqlite for durable local state and telemetry
/dp/frankentui for optional human TUI

Product Shape

The project must be both:

A reusable Rust library for embedding ASR/diarization pipelines.
A standalone CLI binary with:
- robot mode (agent-first, structured machine output)
- optional TUI mode for humans

Input requirements:

Any common audio file format (automatic ffmpeg normalization)
Microphone or line-in capture where supported
Optional low-bandwidth TTY/PTY streaming mode (see architecture docs)

External orchestrated tools (invoked as subprocesses, never vendored): whisper-cli / insanely-fast-whisper / python3 (backends), ffmpeg/ffprobe (normalization fallback), and yt-dlp for the youtube subcommand — each user-installed and overridable via env (FRANKEN_WHISPER_*_BIN; yt-dlp via FRANKEN_WHISPER_YTDLP_BIN).

Porting Workflow (Spec-First)

Use this workflow order:

PLAN_TO_PORT_WHISPER_STACK_TO_RUST.md
EXISTING_LEGACY_WHISPER_STRUCTURE.md
PROPOSED_ARCHITECTURE.md
FEATURE_PARITY.md

Implementation should follow spec documents, not ad-hoc copying from legacy code.

SQLite + JSONL Workflow

Source of truth: SQLite.

JSONL is for auditability, git-friendly backup, and recovery.

Hard rules:

One-way sync at a time (locked + atomic).
No two-way magic merge.
Maintain explicit version markers.
Keep recovery procedures documented in:
- SYNC_STRATEGY.md
- RECOVERY_RUNBOOK.md

Alien-Artifact Engineering Contract

For runtime/adaptive decisions, include:

explicit state space
explicit actions
loss matrix
posterior/confidence terms
calibration metric
deterministic fallback trigger
evidence ledger artifact

No adaptive controller should ship without conservative fallback.

Code Editing Discipline

No Script-Based Changes

NEVER run a script that processes/changes code files in this repo. Brittle regex-based transformations create far more problems than they solve.

Always make code changes manually, even when there are many instances
For many simple changes: use parallel subagents
For subtle/complex changes: do them methodically yourself

No File Proliferation

If you want to change something or add a feature, revise existing code files in place.

NEVER create variations like:

mainV2.rs
main_improved.rs
main_enhanced.rs

New files are reserved for genuinely new functionality that makes zero sense to include in any existing file. The bar for creating new files is incredibly high.

Backwards Compatibility

We do not care about backwards compatibility—we're in early development with no users. We want to do things the RIGHT way with NO TECH DEBT.

Never create "compatibility shims"
Never create wrapper functions for deprecated APIs
Just fix the code directly

Toolchain

Rust 2024 edition.
Nightly toolchain (rust-toolchain.toml).
#![forbid(unsafe_code)].
Cargo only.

Mandatory Checks After Substantive Changes

cargo fmt --check
cargo check --all-targets
cargo clippy --all-targets -- -D warnings
cargo test

If any check fails, fix root causes before handing off.

Testing Policy

Each module should include unit tests for:

happy path
edge cases
error handling

Integration/e2e tests should cover:

file-based transcription flow
robot mode JSON/NDJSON contracts
SQLite + JSONL sync/recovery
ffmpeg normalization behavior and failure paths

Agent Ergonomics Requirements

Robot mode should be:

stable schema
deterministic where possible
explicit error codes
line-oriented output (jsonl/ndjson)
easy to pipe into other tools

Do not mix human decoration with machine output in robot mode.

Multi-Agent / Beads / BV

If using br / bv:

Use br for issue CRUD.
Use only bv --robot-* variants (never bare bv).

Session Completion

Before finishing:

Summarize what changed.
List quality gates run and results.
Document remaining risks/gaps.
Propose concrete next steps.

MCP Agent Mail — Multi-Agent Coordination

A mail-like layer that lets coding agents coordinate asynchronously via MCP tools and resources. Provides identities, inbox/outbox, searchable threads, and advisory file reservations with human-auditable artifacts in Git.

Why It's Useful

Prevents conflicts: Explicit file reservations (leases) for files/globs
Token-efficient: Messages stored in per-project archive, not in context
Quick reads: resource://inbox/..., resource://thread/...

Same Repository Workflow

Register identity:

ensure_project(project_key=<abs-path>)
register_agent(project_key, program, model)

Reserve files before editing:

file_reservation_paths(project_key, agent_name, ["src/**"], ttl_seconds=3600, exclusive=true)

Communicate with threads:

send_message(..., thread_id="FEAT-123")
fetch_inbox(project_key, agent_name)
acknowledge_message(project_key, agent_name, message_id)

Quick reads:

resource://inbox/{Agent}?project=<abs-path>&limit=20
resource://thread/{id}?project=<abs-path>&include_bodies=true

Macros vs Granular Tools

Prefer macros for speed: macro_start_session, macro_prepare_thread, macro_file_reservation_cycle, macro_contact_handshake
Use granular tools for control: register_agent, file_reservation_paths, send_message, fetch_inbox, acknowledge_message

Common Pitfalls

"from_agent not registered": Always register_agent in the correct project_key first
"FILE_RESERVATION_CONFLICT": Adjust patterns, wait for expiry, or use non-exclusive reservation
Auth errors: If JWT+JWKS enabled, include bearer token with matching kid

Beads (br) — Dependency-Aware Issue Tracking

Beads provides a lightweight, dependency-aware issue database and CLI (br - beads_rust) for selecting "ready work," setting priorities, and tracking status. It complements MCP Agent Mail's messaging and file reservations.

Important: br is non-invasive—it NEVER runs git commands automatically. You must manually commit changes after br sync --flush-only.

Conventions

Single source of truth: Beads for task status/priority/dependencies; Agent Mail for conversation and audit
Shared identifiers: Use Beads issue ID (e.g., br-123) as Mail thread_id and prefix subjects with [br-123]
Reservations: When starting a task, call file_reservation_paths() with the issue ID in reason

Typical Agent Flow

Pick ready work (Beads):

br ready --json  # Choose highest priority, no blockers

Reserve edit surface (Mail):

file_reservation_paths(project_key, agent_name, ["src/**"], ttl_seconds=3600, exclusive=true, reason="br-123")

Announce start (Mail):

send_message(..., thread_id="br-123", subject="[br-123] Start: <title>", ack_required=true)

Work and update: Reply in-thread with progress

Complete and release:

br close 123 --reason "Completed"
br sync --flush-only  # Export to JSONL (no git operations)

release_file_reservations(project_key, agent_name, paths=["src/**"])

Final Mail reply: [br-123] Completed with summary

Mapping Cheat Sheet

Concept	Value
Mail `thread_id`	`br-###`
Mail subject	`[br-###] ...`
File reservation `reason`	`br-###`
Commit messages	Include `br-###` for traceability

bv — Graph-Aware Triage Engine

bv is a graph-aware triage engine for Beads projects (.beads/beads.jsonl). It computes PageRank, betweenness, critical path, cycles, HITS, eigenvector, and k-core metrics deterministically.

Scope boundary: bv handles what to work on (triage, priority, planning). For agent-to-agent coordination (messaging, work claiming, file reservations), use MCP Agent Mail.

CRITICAL: Use ONLY --robot-* flags. Bare bv launches an interactive TUI that blocks your session.

The Workflow: Start With Triage

bv --robot-triage is your single entry point. It returns:

quick_ref: at-a-glance counts + top 3 picks
recommendations: ranked actionable items with scores, reasons, unblock info
quick_wins: low-effort high-impact items
blockers_to_clear: items that unblock the most downstream work
project_health: status/type/priority distributions, graph metrics
commands: copy-paste shell commands for next steps

bv --robot-triage        # THE MEGA-COMMAND: start here
bv --robot-next          # Minimal: just the single top pick + claim command

Command Reference

Planning:

Command	Returns
`--robot-plan`	Parallel execution tracks with `unblocks` lists
`--robot-priority`	Priority misalignment detection with confidence

Graph Analysis:

Command	Returns
`--robot-insights`	Full metrics: PageRank, betweenness, HITS, eigenvector, critical path, cycles, k-core, articulation points, slack
`--robot-label-health`	Per-label health: `health_level`, `velocity_score`, `staleness`, `blocked_count`
`--robot-label-flow`	Cross-label dependency: `flow_matrix`, `dependencies`, `bottleneck_labels`
`--robot-label-attention [--attention-limit=N]`	Attention-ranked labels

History & Change Tracking:

Command	Returns
`--robot-history`	Bead-to-commit correlations
`--robot-diff --diff-since <ref>`	Changes since ref: new/closed/modified issues, cycles

Other:

Command	Returns
`--robot-burndown <sprint>`	Sprint burndown, scope changes, at-risk items
`--robot-forecast <id\|all>`	ETA predictions with dependency-aware scheduling
`--robot-alerts`	Stale issues, blocking cascades, priority mismatches
`--robot-suggest`	Hygiene: duplicates, missing deps, label suggestions
`--robot-graph [--graph-format=json\|dot\|mermaid]`	Dependency graph export
`--export-graph <file.html>`	Interactive HTML visualization

Scoping & Filtering

bv --robot-plan --label backend              # Scope to label's subgraph
bv --robot-insights --as-of HEAD~30          # Historical point-in-time
bv --recipe actionable --robot-plan          # Pre-filter: ready to work
bv --recipe high-impact --robot-triage       # Pre-filter: top PageRank
bv --robot-triage --robot-triage-by-track    # Group by parallel work streams
bv --robot-triage --robot-triage-by-label    # Group by domain

Understanding Robot Output

All robot JSON includes:

data_hash — Fingerprint of source beads.jsonl
status — Per-metric state: computed|approx|timeout|skipped + elapsed ms
as_of / as_of_commit — Present when using --as-of

Two-phase analysis:

Phase 1 (instant): degree, topo sort, density
Phase 2 (async, 500ms timeout): PageRank, betweenness, HITS, eigenvector, cycles

jq Quick Reference

bv --robot-triage | jq '.quick_ref'                        # At-a-glance summary
bv --robot-triage | jq '.recommendations[0]'               # Top recommendation
bv --robot-plan | jq '.plan.summary.highest_impact'        # Best unblock target
bv --robot-insights | jq '.status'                         # Check metric readiness
bv --robot-insights | jq '.Cycles'                         # Circular deps (must fix!)

UBS — Ultimate Bug Scanner

Golden Rule: ubs <changed-files> before every commit. Exit 0 = safe. Exit >0 = fix & re-run.

Commands

ubs file.rs file2.rs                    # Specific files (< 1s) — USE THIS
ubs $(git diff --name-only --cached)    # Staged files — before commit
ubs --only=rust,toml src/               # Language filter (3-5x faster)
ubs --ci --fail-on-warning .            # CI mode — before PR
ubs .                                   # Whole project (ignores target/, Cargo.lock)

Output Format

⚠️  Category (N errors)
    file.rs:42:5 – Issue description
    💡 Suggested fix
Exit code: 1

Parse: file:line:col → location | 💡 → how to fix | Exit 0/1 → pass/fail

Fix Workflow

Read finding → category + fix suggestion
Navigate file:line:col → view context
Verify real issue (not false positive)
Fix root cause (not symptom)
Re-run ubs <file> → exit 0
Commit

Bug Severity

Critical (always fix): Memory safety, use-after-free, data races, SQL injection
Important (production): Unwrap panics, resource leaks, overflow checks
Contextual (judgment): TODO/FIXME, println! debugging

RCH — Remote Compilation Helper

RCH offloads cargo build, cargo test, cargo clippy, and other compilation commands to a fleet of 8 remote Contabo VPS workers instead of building locally. This prevents compilation storms from overwhelming csd when many agents run simultaneously.

RCH is installed at ~/.local/bin/rch and is hooked into Claude Code's PreToolUse automatically. Most of the time you don't need to do anything if you are Claude Code — builds are intercepted and offloaded transparently.

To manually offload a build:

rch exec -- cargo build --release
rch exec -- cargo test
rch exec -- cargo clippy

Quick commands:

rch doctor                    # Health check
rch workers probe --all       # Test connectivity to all 8 workers
rch status                    # Overview of current state
rch queue                     # See active/waiting builds

If rch or its workers are unavailable, it fails open — builds run locally as normal.

Note for Codex/GPT-5.2: Codex does not have the automatic PreToolUse hook, but you can (and should) still manually offload compute-intensive compilation commands using rch exec -- <command>. This avoids local resource contention when multiple agents are building simultaneously.

ast-grep vs ripgrep

Use ast-grep when structure matters. It parses code and matches AST nodes, ignoring comments/strings, and can safely rewrite code.

Refactors/codemods: rename APIs, change import forms
Policy checks: enforce patterns across a repo
Editor/automation: LSP mode, --json output

Use ripgrep when text is enough. Fastest way to grep literals/regex.

Recon: find strings, TODOs, log lines, config values
Pre-filter: narrow candidate files before ast-grep

Rule of Thumb

Need correctness or applying changes → ast-grep
Need raw speed or hunting text → rg
Often combine: rg to shortlist files, then ast-grep to match/modify

Rust Examples

# Find structured code (ignores comments)
ast-grep run -l Rust -p 'fn $NAME($$ARGS) -> $RET { $$BODY }'

# Find all unwrap() calls
ast-grep run -l Rust -p '$EXPR.unwrap()'

# Quick textual hunt
rg -n 'println!' -t rust

# Combine speed + precision
rg -l -t rust 'unwrap\(' | xargs ast-grep run -l Rust -p '$X.unwrap()' --json

Morph Warp Grep — AI-Powered Code Search

Use mcp__morph-mcp__warp_grep for exploratory "how does X work?" questions. An AI agent expands your query, greps the codebase, reads relevant files, and returns precise line ranges with full context.

Use ripgrep for targeted searches. When you know exactly what you're looking for.

Use ast-grep for structural patterns. When you need AST precision for matching/rewriting.

When to Use What

Scenario	Tool	Why
"How does the cancellation protocol work?"	`warp_grep`	Exploratory; don't know where to start
"Where is the region close logic?"	`warp_grep`	Need to understand architecture
"Find all uses of `Cx::trace`"	`ripgrep`	Targeted literal search
"Find files with `println!`"	`ripgrep`	Simple pattern
"Replace all `unwrap()` with `expect()`"	`ast-grep`	Structural refactor

warp_grep Usage

mcp__morph-mcp__warp_grep(
  repoPath: "/data/projects/franken_whisper",
  query: "How does the structured concurrency scope API work?"
)

Returns structured results with file paths, line ranges, and extracted code snippets.

Anti-Patterns

Don't use warp_grep to find a specific function name → use ripgrep
Don't use ripgrep to understand "how does X work" → wastes time with manual reads
Don't use ripgrep for codemods → risks collateral edits

Beads Workflow Integration

This project uses beads_rust (br) for issue tracking. Issues are stored in .beads/ and tracked in git.

Important: br is non-invasive—it NEVER executes git commands. After br sync --flush-only, you must manually run git add .beads/ && git commit.

Essential Commands

# View issues (launches TUI - avoid in automated sessions)
bv

# CLI commands for agents (use these instead)
br ready              # Show issues ready to work (no blockers)
br list --status=open # All open issues
br show <id>          # Full issue details with dependencies
br create --title="..." --type=task --priority=2
br update <id> --status=in_progress
br close <id> --reason "Completed"
br close <id1> <id2>  # Close multiple issues at once
br sync --flush-only  # Export to JSONL (NO git operations)

Workflow Pattern

Start: Run br ready to find actionable work
Claim: Use br update <id> --status=in_progress
Work: Implement the task
Complete: Use br close <id>
Sync: Run br sync --flush-only then manually commit

Key Concepts

Dependencies: Issues can block other issues. br ready shows only unblocked work.
Priority: P0=critical, P1=high, P2=medium, P3=low, P4=backlog (use numbers, not words)
Types: task, bug, feature, epic, question, docs
Blocking: br dep add <issue> <depends-on> to add dependencies

Session Protocol

Before ending any session, run this checklist:

git status              # Check what changed
git add <files>         # Stage code changes
br sync --flush-only    # Export beads to JSONL
git add .beads/         # Stage beads changes
git commit -m "..."     # Commit everything together
git push                # Push to remote

Best Practices

Check br ready at session start to find available work
Update status as you work (in_progress → closed)
Create new issues with br create when you discover tasks
Use descriptive titles and set appropriate priority/type
Always br sync --flush-only && git add .beads/ before ending session

Landing the Plane (Session Completion)

When ending a work session, you MUST complete ALL steps below.

MANDATORY WORKFLOW:

File issues for remaining work - Create issues for anything that needs follow-up
Run quality gates (if code changed) - Tests, linters, builds
Update issue status - Close finished work, update in-progress items
Sync beads - br sync --flush-only to export to JSONL
Hand off - Provide context for next session

cass — Cross-Agent Session Search

cass indexes prior agent conversations (Claude Code, Codex, Cursor, Gemini, ChatGPT, etc.) so we can reuse solved problems.

Rules: Never run bare cass (TUI). Always use --robot or --json.

Examples

cass health
cass search "async runtime" --robot --limit 5
cass view /path/to/session.jsonl -n 42 --json
cass expand /path/to/session.jsonl -n 42 -C 3 --json
cass capabilities --json
cass robot-docs guide

Tips

Use --fields minimal for lean output
Filter by agent with --agent
Use --days N to limit to recent history

stdout is data-only, stderr is diagnostics; exit code 0 means success.

Treat cass as a way to avoid re-solving problems other agents already handled.

Note for Codex/GPT-5.2:

You constantly bother me and stop working with concerned questions that look similar to this:

Unexpected changes (need guidance)

- Working tree still shows edits I did not make in Cargo.toml, Cargo.lock, src/runtime.rs, src/scope.rs. Please advise whether to keep/commit/revert these before any further work. I did not touch them.

Next steps (pick one)

1. Decide how to handle the unrelated modified files above so we can resume cleanly.

NEVER EVER DO THAT AGAIN. The answer is literally ALWAYS the same: those are changes created by the potentially dozen of other agents working on the project at the same time. This is not only a common occurence, it happens multiple times PER MINUTE. The way to deal with it is simple: you NEVER, under ANY CIRCUMSTANCE, stash, revert, overwrite, or otherwise disturb in ANY way the work of other agents. Just treat those changes identically to changes that you yourself made. Just fool yourself into thinking YOU made the changes and simply don't recall it for some reason.

Note on Built-in TODO Functionality

Also, if I ask you to explicitly use your built-in TODO functionality, don't complain about this and say you need to use beads. You can use built-in TODOs if I tell you specifically to do so. Always comply with such orders.

FilesExpand file tree

AGENTS.md

Latest commit

History