Skip to content

CHANGELOG.md

Latest commit

 

History

History
258 lines (219 loc) · 15.3 KB

CHANGELOG.md

File metadata and controls

258 lines (219 loc) · 15.3 KB

Changelog

All notable changes to this project are documented here. Format based on Keep a Changelog; this project adheres to Semantic Versioning.

[Unreleased]

Added

  • Independent out-of-band re-execution watcher (prd_taskmaster/tournament/watcher.py; watcher-run / watcher-status CLI) — the precondition for ever enabling real (--enforce-slash) forfeiture. It re-adjudicates settled tournament submissions from primary evidence (the claimed commit + the CDD card) without trusting the recorded verdict: it re-runs the oracle gate, re-derives sha256(diff base..HEAD) to catch diff-copy tamper independently of the live collector, and accumulates a concordance ledger over real slash decisions. A permit_enforce_slash gate is fail-closed: real slashing is permitted only when every to-be-slashed submission is independently confirmed, there is no discrepancy or abstain anywhere in the job, and the watcher's historical concordance clears a threshold over a minimum number of prior decisions. Inability to verify (oracle could not run, no worktree, failed hash recompute) abstains — it is never counted as grounds to slash.
  • Engine-enforced shadow-until-permitted — run_tournament now consults the watcher before any real --enforce-slash settle and downgrades to shadow unless the watcher permits the job (fail-closed on any watcher error). Real AtlasCoin is never burned without an independent positive confirmation; the shadow-slash default path is unchanged.

[5.3.0] — 2026-06-17

The "unfakable done" release: a task ships only when its test genuinely re-ran green AND the code it tested is wired in — and that trustworthy signal now drives a competitive marketplace. Consolidates the oracle + reachability gates, the settled tournament, and the dogfooded first-run UX into one engine.

Added

  • Re-execution oracle (ship-check Gate 5) — re-grades each "done" task by re-running the operator-held tests at the claimed commit in a network-isolated, digest-pinned podman sandbox. A submitter can no longer pass by editing their own logs; the self-grantable SHIP_CHECK_OVERRIDE_ADMIN bypass is removed.
  • Reachability gate (Gate 6) — done = oracle-PASS AND the tested code is wired in. A green test on an orphan module (imported by nothing) is blocked and surfaced as ⚠ scaffolded.
  • Settled-tournament marketplace (tournament run / tournament-status) — N executors race one job, every submission is adjudicated through both gates, the winner is paid in AtlasCoin, honest losers are refunded, and a trusted reputation store (UCB explore/exploit) routes the next job to the cheapest proven-capable executor. Includes a cheap OpenRouter (goose) racer.
  • Deterministic expand-structural — decomposes an under-specified task into ≥2 verifiable subtasks with no model or network.

Changed

  • /atlas now opens with a Confirm-Intent / plan-mode gate before any file is written, and gate prompts are harness-adaptive (AskUserQuestion on Claude Code; nearest equivalent on codex/gemini).
  • task-master-ai is now optional (the native engine stands alone); the tm-parallel/tm-plan/tm-run/tm-harvest surfaces were removed.
  • Fakery is shadow-logged only this release — proven cheating is recorded, no AtlasCoin is burned, and honest losers always get their stake back.

Fixed

  • Preflight configure no longer silently no-ops on a fresh project (returns an explicit "configuration deferred").
  • tournament settle always returns a parseable {ok, stage} envelope (no empty stdout on a bad job dir); settlement is crash-resumable (no paid-but-stuck winner).
  • The cheap (goose) racer is no longer falsely rejected on non-ASCII diffs; tournament racers are wired to the real orchestrator inbox.
  • AtlasCoin is conserved (no mint/burn) under account aliasing.

Tests

  • 872 passing + 4 real-podman e2e gates green (oracle dogfood + tournament settle/pay/reputation).

[5.2.2] — 2026-06-14

Front-door UX flow fixes (a UX-flow audit found the journey still broke before a new user reached the 5.2.1 backend fixes). See docs/audit/UX-FLOW-AUDIT.md. Also syncs the version source-of-truth (prd_taskmaster/__init__.py), which 5.2.1 missed.

Fixed

  • UX-P0-1 — the README's first command now resolves. README led with /atlas, which a fresh /plugin install prd does not provide (plugin commands are namespaced /prd:*). Added a brand-name atlas entrypoint skill (→ /prd:atlas, a thin alias that dispatches to the go orchestrator) and updated the README first-run to /prd:atlas (or /prd:go, or natural language).
  • UX-P0-2 — phase gates no longer document a self-contradicting STOP. setup/ discover/generate/handoff opened with "if the gate fails, stop" immediately followed by "it WILL fail on first entry, proceed past it (see morning brief)" — a compliant autonomous agent would halt. Rewritten to explain check_gate is an EXIT gate (evidence to advance, not to enter), so a first-entry false is expected; the gate is enforced on advance. Removed leaked internal references ("morning brief", "Mum dogfood feedback").
  • UX-P0-3 — token_economy set via /customise-workflow is now honored. It writes .atlas-ai/config/atlas.json, but the engine read economy only from .atlas-ai/fleet.json. load_fleet_config now reads token_economy from atlas.json when fleet.json doesn't set one (fleet.json wins if it does); the config schema doc adds the key.

Fixed (version hygiene)

  • prd_taskmaster/__init__.py version bumped (5.2.1 set package.json/plugin.json but missed the __init__.py source-of-truth the manifest tests check).

[5.2.1] — 2026-06-14

Pre-relaunch hardening — fixes the first-run failures a multi-agent audit found on the documented zero-key path (and reproduced firsthand). See docs/audit/.

Fixed

  • P0-1 — configure-providers now REPAIRS keyless stock defaults. It previously only filled empty model roles, so task-master init's paid anthropic+perplexity defaults survived untouched and a keyless first run produced 0 tasks. It now migrates any stock default whose provider is unusable in the current environment to the available claude/codex CLI or the free local proxy (KNOWN_STOCK_TASKMASTER_DEFAULTS
    • a _provider_usable check). Genuine user configs and usable configs are preserved; the provider decision now dominates the tier decision.
  • P0-2 — the SETUP gate (validate_setup) is credential-aware. It reported ready=True whenever a model-id string was present — green-lighting the exact 0-task config. It now verifies the configured provider is actually reachable (key present / CLI on PATH).
  • P0-3 — expand degrades to a structural pass when the research provider is down. Both the parallel and serial paths hardcoded --research; a quota/auth outage hard-failed tasks to 0 subtasks. They now retry without --research (structural expand is always available) and mark the result degraded.
  • P1-1 — parse_prd reports ok=False on a 0-task parse (was a silent success).
  • Nested-session spawn probe (gh #11/#12). When main is a CLI-spawning provider inside a nested Claude Code session, the gate now probes whether the spawn works rather than assuming — keeping the free path when it works and surfacing an actionable error (never bare --claude-code) when it genuinely fails.
  • Stale-tag detection (gh #13). preflight now reports current_tag_stale + suggested_fresh_tag so a new PRD is not parsed into a polluted, fully-done tag.

Changed

  • package.json now carries author, homepage, and bugs so the npm page is not anonymous.

[5.2.0] — 2026-06-13

Progress visualization — the README/UX-SPEC panels are now real output.

Added

  • status command + render_status MCP tool that draw the boxed Atlas progress panels from real pipeline state: phase tracker, the GENERATE validation scorecard (grade bar + checks + line-located warnings + placeholders + task/subtask counts), the preflight capability panel, the handoff panel, the execute progress bar, and the ship-check gates panel. --format boxed|ascii|json, --all, --phase.
  • prd_taskmaster/render.py (pure renderer, UX-SPEC §7 symbol grammar with ASCII fallbacks via ATLAS_ASCII=1; display-width-aware so borders align even with the 🔒 emoji) and prd_taskmaster/status.py (pure reader of pipeline/tasks/validation state).
  • validate-prd now persists its result to .atlas-ai/state/validation.json so the scorecard renders without recomputation.
  • Phase skills (setup, generate, handoff, execute-task) render the matching panel at each phase boundary.

Changed

  • The grade bar floors rather than rounds (49/57 = 86% → 8/10 filled), matching the README mockup and never inflating a grade.
  • UX-SPEC pricing mockups marked superseded by the private-pilot decision (no $29).

[5.1.2] — 2026-06-13

Honesty + positioning + discoverability.

Added

  • Pre-alpha status is now stated up front (badge + Project status section): the newer systems (fleet, backend abstraction, token economy, Pro MCPs) are not yet battle-tested; expect breaking changes.
  • PRD validation warnings (vague language + placeholders) now carry line numbers — the "quoted + located, not just counted" claim is now executable truth, not just rendering.
  • package.json keywords for npm search discoverability (factual token-economy / cost-ledger / model-routing mechanism terms — no unproven savings claims).

Changed

  • Atlas Pro reframed as a private pilot. Not generally available, not for sale; pricing removed. Access is granted at discretion during the pilot; "Get Atlas Pro" CTAs become "Request pilot access" pointing at GitHub Discussions (an atlas-ai.au/pilot signup page is in progress).

[5.1.1] — 2026-06-13

Audit-driven honesty release (dogfood cycle 6 — the engine ran its own pipeline on this work).

Added

  • Placeholder hard fail. Any placeholder — {{...}}, bracketed, or bare case-sensitive TBD/TODO — now floors the PRD grade to NEEDS_WORK, sets hard_fail in the result, and makes validate-prd exit non-zero. The README's rigor claim is now executable truth.
  • Installer pins its clone to its own release tag (--branch v$VERSION, branch fallback).

Fixed

  • atlas-ai.au/install and atlas-ai.au/pro now resolve (Cloudflare single-redirects) — the recommended install path 404'd at 5.1.0 launch.
  • README discloses the audit-logged ship-check admin override, the npm postinstall pip step, and marks the local research proxy bring-your-own.
  • Marketplace manifest description no longer reads as an internal testing artifact.

[5.1.0] — 2026-06-13

First npm publish (prd-taskmaster@5.1.0). This release consolidates the two development lines into one artifact and re-launches the repo as an open-core product: a free MIT engine plus a commercial Atlas Pro tier.

Lineage: the public v4.x line (this repo) absorbed the private v5.x plugin line (prd-taskmaster-plugin, internal name "atlas-go", final state v5-final @ f140490) via file-level imports marked with Imported-From: commit trailers — state machine, ship-check gates, execute-task hardening, npm pack hygiene, and the granular validation test suite. Versioning continues from the higher v5.x line so the consolidated artifact supersedes both.

Added (5.1.0 consolidation)

  • Plugin namespace renamed prd-taskmaster → prd — commands are now /prd:go etc.; MCP tool ids are mcp__plugin_prd_go__* (old prefixes kept as legacy allowed-tools aliases).
  • Per-directory .npmignore files keep Python bytecode out of the npm tarball (npm 11 ignores the root .npmignore inside files[]-allowlisted directories).
  • 21 granular PRD-validation tests ported from the plugin suite (tests/core/test_validation.py).
  • Setup/execute-task skills resolve the customizations starter pack and ship-check.py from the packaged ${CLAUDE_PLUGIN_ROOT}/skel/ (previously referenced a developer-machine path).

The notes below document the v4.0.0 line that this release ships for the first time.

Added

  • Token economy (token_economy: conservative|balanced|performance) — per-op-class start tiers, validator-gated escalation with per-mode ceilings, and economy-aware provider configuration. Verified priors and sources in docs/product/MODEL-ECONOMY.md.
  • Parallel native TaskMaster expansion (tm-parallel / tm_parallel_expand) — TaskMaster's model-agnostic expand --research runs concurrently in isolated workdirs (per-task economy-tier models), merged atomically; the Claude-subagent path becomes the documented fallback.
  • Local cost telemetry (.atlas-ai/telemetry.jsonl) + economy-report summarizer.
  • Unified deterministic core (prd_taskmaster/) — a single stdlib-only Python package that is the one source of truth for PRD validation, task calculation, complexity enrichment, capability detection, the pipeline state machine, and the ship-check gate. Imported by both the zero-dependency skill (script.py) and the FastMCP plugin (mcp-server/server.py).
  • 5-phase gated pipeline — SETUP → DISCOVER → GENERATE → HANDOFF → EXECUTE with atomic compare-and-swap transitions over a flock-guarded pipeline.json.
  • CDD execute loop — skills/execute-task runs a 13-step contract-driven cycle with evidence cards; completion is gated by a deterministic SHIP_CHECK_OK token (a non-zero exit code in any evidence file blocks it).
  • Parallel research fan-out (prd_taskmaster/parallel.py) — plan/apply research packets let an agent expand tasks across parallel subagents and merge results atomically.
  • /atlas command — the primary user-facing invocation (alias of the orchestrator skill).
  • Product + UX specs (docs/product/) — the living contract the dogfood ship-check verifies against.
  • Atlas Pro teaser — the handoff surfaces Atlas Fleet (parallel multi-session execution) as a locked, clearly-priced upgrade. The free engine stays fully functional standalone.

Changed

  • Repo keeps the name prd-taskmaster (and its 508★). The product/command brand is Atlas; the internal "atlas-go" name from the plugin line is retired.
  • PRD validation is stricter than older internal versions (grade thresholds + placeholder attribution). PRDs that previously passed loosely may now score lower — by design.
  • Two install paths from one repo: curl one-liner (zero-dependency skill) and Claude Code plugin / npm install (full FastMCP plugin).

Fixed

  • Ship-check tests aligned to the live skel/ship-check.py gate contract (3 stale-schema failures resolved).

Lineage

v4.0.0 merges the prd-taskmaster-v2 skill line (internally versioned 2.x, never published) and the atlas-go plugin line (internally 5.x, never published) back into this repository. Those internal version numbers do not appear as tags here; the public lineage runs v3.0.0 → v4.0.0.

[3.0.0] — 2026-02-12

  • Codified deterministic operations into script.py; curl installer with update notifications; template-based PRD and CLAUDE.md generation; community files. (Pre-merge single-skill product.)