| name | tackle-todos |
| description | Read TODO.md and scan source code for TODO/FIXME comments, build a consolidated list, then dispatch parallel agents to address as many items as possible. |
| user-invocable | true |
| allowed-tools | Bash, Read, Write, Edit, Grep, Glob, Agent |
Tackle TODOs: Consolidate and Resolve Open Items
This skill scans the project for all outstanding work items — both from .agents/docs/TODO.md and from // TODO / // FIXME comments in source code — builds a consolidated, deduplicated, prioritized list, and then dispatches parallel agents to resolve as many items as possible.
Use this skill when: you want to make a focused sweep of outstanding TODOs and fix them in bulk.
Arguments
[filter] (optional): A keyword to restrict which TODOs to tackle (e.g., ingest, stream, ffi, error). If omitted, all TODOs are considered.
Step 0: Collect TODOs from TODO.md
Read .agents/docs/TODO.md and extract every unchecked - [ ] item. Record its area, description, and source.
Step 1: Scan Source Code for TODO/FIXME Comments
Use Grep to search the repo for // TODO / TODO: / FIXME in src/**/*.rs (the extern "C" surface) and in the C/C++ examples and headers (examples/**/*.c, examples/**/*.cpp, include/*.h, include/*.hpp).
Group and deduplicate the results:
- Identical or near-identical comments repeated across many files should be collapsed into a single work item with a note about which files/packages are affected.
- Comments that are informational-only (e.g., noting a known limitation that cannot be fixed without large design changes) should be flagged but deprioritized.
Step 2: Build a Consolidated TODO List
Merge the two sources into a single list. Deduplicate items that appear in both TODO.md and as code comments.
For each item, assign a category:
| Category | Description | Priority |
|---|
| systematic | Same pattern repeated across many files | High — one fix propagates widely |
| behavioural | Missing or incorrect behaviour affecting correctness | High |
| validation | Missing input validation or error responses | Medium |
| serialization | Wire format, encoding, or protocol bugs | Medium |
| test-only | TODO in test code noting a test gap, not a code defect | Low |
| design | Requires significant design work or new abstractions | Deferred — flag for user |
Write the consolidated list to .agents-workspace/tmp/consolidated-todos.md for reference.
Step 2b: Verify stale items before dispatch
Before dispatching agents on any TODO entry that is more than 24 hours old, verify the entry is still applicable: grep the relevant code for the symptom (the call site, the missing handler, etc.). Skip or close stale entries instead of dispatching.
Step 3: Filter (if argument provided)
If the user passed a [filter] argument, restrict the work list to items matching that filter (package name or keyword).
Step 4: Plan Parallel Work Items
Group the consolidated TODOs into independent, parallelizable work units. Each work unit should:
- Be self-contained (touching one package or one cross-cutting concern)
- Not conflict with other parallel work units (no two agents editing the same file)
Present the plan to the user and get confirmation before dispatching.
Step 5: Dispatch Parallel Agents
For each approved work unit, launch an Agent (subagent_type: general-purpose) with a clear prompt that includes:
- The specific TODO(s) to address
- The files to modify
- The expected behaviour
- Instructions to run the local gate after making changes (per
AGENTS.md): cargo fmt --check, cargo build --release (this also regenerates include/imbh.h via cbindgen), cargo clippy --all-targets --release -- -D warnings, and cargo test --release. If the ABI, header, or examples changed, also cmake -S . -B build -DCMAKE_BUILD_TYPE=Release && cmake --build build && ctest --test-dir build --output-on-failure (the examples link target/release/libimbh_c.{so,a}, so build the crate first).
Launch as many agents in parallel as there are independent work units, in a single batch.
❌ Never use isolation: worktree for parallel agents — it has repeatedly caused trouble. Instead, launch tasks that are independent of one another in a batch.
Step 6: Collect Results and Update TODO.md
After all agents complete:
- Review each agent's results — check if the change built, vetted, and tests passed.
- For successfully resolved items, mark them as
- [x] in .agents/docs/TODO.md and remove the corresponding // TODO comments from source code.
- For items that could not be resolved, add notes about why and what was attempted.
- Append a summary to
.agents/docs/JOURNAL.md documenting what was tackled and the outcomes.
Notes
- Do not attempt design-category items without user approval. These require architectural decisions.
- Respect
AGENTS.md rules: no git checkout, no git restore, no discretionary commits. Agents should only edit files, not commit.
- Always run
cargo test --release (the tests/roundtrip.rs suite drives the extern "C" surface in-process) and, when the ABI/header/examples changed, the CMake/CTest examples. FFI changes must keep the ownership invariants passing — an empty result still carries its typed schema, and the ArrowArrayStream stays valid after the DB reclaims segments. Keep include/imbh.h freshly regenerated by a cargo build.