task-backlog

name	task-backlog
description	Task-backlog lifecycle tooling for the Epistemic-Humility-Research repo - one markdown file per task under backlog/{tasks,drafts,completed}, a bin/task CLI (new/list/show/claim/release/review/done/validate), a generated TODO.md table, and a pre-commit commit gate that requires gated work files to be covered by an active task. Use to mint, claim, and close backlog tasks, to bind a task to an experiments/<slug>/ record, and to understand the commit-gate scope and escape hatch. Tasks point at experiments; experiments keep their own lifecycle under bin/exp and AMENDMENT.md.
allowed-tools	Read, Bash, Write, Grep, Glob

Task Backlog

Tasks live in backlog/ -- one markdown file per task. All task operations go through bin/task -- it mints merge-safe random hex ids (never sequential; those collide when parallel worktrees merge), enforces the schema, and re-validates after every mutating command. This ports the syntunia backlog/ task-lifecycle pattern into this repo, adapted to bind tasks against this repo's experiments/<slug>/ records instead of a generic component: path.

Experiments themselves stay governed by bin/exp and experiments/<slug>/AMENDMENT.md (see the experiments skill). Tasks never duplicate an experiment's lifecycle -- they only point at one via the optional experiment: field, and validation fails a task that stays open against an experiment that has already reached a terminal status (the "rot-killer").

States (enumerated -- the validator rejects anything else)

draft (in backlog/drafts/) -> todo -> in-progress -> in-review -> done (archived to backlog/completed/)

The CLI

python3 bin/task.py new "Title" --tier P --priority high \
    [--experiment <slug>] [--component path] \
    [--depends-on task-xxxxxx]... [--file existing/path]... \
    [--new-file planned/path]... [--blocker "text"] [--draft]
python3 bin/task.py list                          # blocked tasks are flagged
python3 bin/task.py show task-xxxxxx               # full working context
python3 bin/task.py claim task-xxxxxx --as @you    # -> in-progress
python3 bin/task.py release task-xxxxxx            # -> todo, unassigned
python3 bin/task.py review task-xxxxxx             # -> in-review (PR open)
python3 bin/task.py done task-xxxxxx               # -> done, archived
python3 bin/task.py validate                       # full validation pass

(bin/task -- no extension -- is the same wrapper shape as bin/exp; either invocation works. bin/task runs the generated .agents/skills/task-backlog/ mirror, so re-sync after editing canonical scripts -- see Maintenance below.)

Rules

1. All work starts from a task. No task covering your change? bin/task new first (or claim an existing one).
2. Claim before you touch a gated file (claim --as @you; default handle is @agent if omitted). Claim refuses a task another assignee holds.
3. Dependencies are a DAG. --depends-on (or a hand-edited depends_on: list) is validated for cycles and unknown ids.
4. files: / new_files: give the change's scope. files: = existing files in scope (validated to exist); new_files: = files the task expects to create (no existence check -- a trailing / marks a whole subtree prefix, honored by the commit gate). Move an entry from new_files: to files: once it exists.
5. experiment: binds a task to experiments/<slug>/ (must exist). validate fails any open task (draft/todo/in-progress/in-review) whose bound experiment's experiment.yaml status: is terminal (resolved, null-result, falsified, historical, shelved) -- close the task (done or release) instead of leaving it to rot. Non-terminal statuses (draft, signed, running) pass.
6. component: is an optional path-prefix binding (. = repo-wide); must resolve to a real path if set.
7. Never hand-edit TODO.md's generated task-backlog block -- see Generated TODO.md block below.

Task file schema

id: task-a1b2c3          # random 6-hex, minted ONLY by the CLI
title: ...
status: todo             # draft | todo | in-progress | in-review | done
assignee: []             # ['@handle']
tier: P                  # A | L | P  (A=amendment-linked, L=lab-notebook, P=paper/infra)
priority: medium         # high | medium | low
experiment: <slug>       # optional; must exist under experiments/
component: papers/paper-5-actuation   # optional path prefix; must exist ('.' = repo-wide)
depends_on: []           # task ids, DAG-validated
files: []                # existing files in scope (existence validated)
new_files: []            # planned files; trailing "/" = subtree prefix
blocker: ""              # optional free text
created_date: YYYY-MM-DD
updated_date: YYYY-MM-DD

Body: ## Description (the why) -- ## Acceptance Criteria (checkboxes) -- ## Work Log (append-only; edit this section by hand as you go).

Generated TODO.md block

build_todo_index.py renders every non-done task into the fenced <!-- BEGIN GENERATED: task-backlog --> / <!-- END GENERATED: task-backlog --> block in TODO.md -- a table (id, title, tier, status, priority, blocker/deps, experiment) sorted by priority then created_date, plus a one-line count of tasks archived in backlog/completed/. Everything outside the fence (the "Prioritized backlog" preamble, dated gotcha paragraphs, the standing-posture note, the "Parked" section, the archive note) is hand-written prose and is preserved verbatim on regen -- never hand-edit inside the fence; run:

python3 .agents/skills/task-backlog/scripts/build_todo_index.py --write
python3 .agents/skills/task-backlog/scripts/build_todo_index.py --check   # CI/pre-commit gate

This is a second, independent generated block in TODO.md alongside the pre-existing <!-- BEGIN GENERATED: amendment-index --> block (from bin/build_backlog_index.py); the two never overlap.

Commit gate (check_task_gate.py)

Wired into .githooks/pre-commit, this enforces change -> task traceability for a fixed gated scope: papers/, bin/, .skills/, .githooks/, .claude/hooks/, and docs/ excluding docs/sessions/. Everything else is exempt, including experiments/, backlog/, TODO.md, docs/sessions/, the generated skill mirrors (.agents/, .claude/skills/), the synaptic-tuner submodule pointer, .claude/settings.json, and analysis/.

A gated file may only be staged when an active task (in-progress or in-review) covers it via files:, new_files: (trailing-slash prefix), or component: prefix, AND that task's own file is also staged (or its updated_date is today, in pre-commit mode). On failure the gate prints exactly which files are uncovered and the one-line fix (bin/task new / bin/task claim).

Escape hatch: EHR_TASK_OK=1 git commit ... skips the gate loudly (it prints that it was skipped) -- use only for genuine emergencies.

Validation

bin/task validate (also run at commit time via .githooks/pre-commit) checks: schema/enum conformance, unique ids, depends_on DAG acyclicity, experiment: slug existence, files: existence, component: existence, and the terminal-experiment rot-killer cross-check above.

Maintenance

Canonical source is .skills/task-backlog/; .agents/skills/task-backlog/ and .claude/skills/task-backlog/ are GENERATED MIRRORS -- never hand-edit a mirror. After editing the canonical scripts or this doc:

python3 bin/sync_skills.py --write --skill task-backlog
python3 bin/sync_skills.py --check --skill task-backlog