speccoder 0.6.1

Changed

• [model] max_run_tokens defaults to 500000 (was 200000), enough for
  an attempt's 30 turns.

Fixed

• list_files and search_code read "." as the workspace instead of
  refusing it (empty path): an agent no longer takes the project for
  empty.
• A spent token budget stops the task with BUDGET-EXHAUSTED and the
  setting to raise, not RUN-FAILED with Exhausted(Recorded { … }).
• A change the task or its acceptance tests need outside the task's
  paths (an entry point, a module they import) is in scope and shown to
  the reviewer; the agent no longer stops for it.

speccoder 0.6.0

Added

• Shape. A writing check as you type (no model) and ✦ Shape,
  which proposes testable words; code marks what it added and dropped,
  and you decide each before the words are used. speccoder shape,
  review edit --shape.
• Needs You by what it means: Blocked on you (oldest first), Ready
  to review (riskiest evidence first, its reasons as chips), Working,
  Idle — in the app and speccoder inbox (--json: groups,
  working, idle, risk).
• Approvals of one tool decided together: Approve N / Reject N…,
  speccoder task approve T1 T2 …; each still its own decision.
• A tick made by an external agent carries Executor: <preset>.

Changed

• Notifications only when something blocks work (a call waiting, a task
  that did not pass), named by the task's title. The sidebar and Dock
  badges count only Blocked on you.
• A parallel lane or bisection worktree starts with the feature
  worktree's target/ and Node folders cloned copy-on-write when the
  lockfile is the same (APFS, Btrfs, XFS): a Rust lane's first build went
  from 6.3 s to 0.2 s. .venv/ is never cloned.
• Built on agentplane 0.56.1.

Fixed

• An OpenAI model sending its structured answer as two messages no longer
  stops the task with ACCEPTANCE-UNWRITTEN (agentplane 0.56.1).
• An author's answer that cannot be used says why, in the provider's
  words, instead of generation did not succeed: failed.

Upgrading

• A finished feature and a stall no longer notify: add "feature_done",
  "stalled" to [ui.notify] on to keep them.
• speccoder replay of a run recorded before this release reports a
  divergence (model-call arguments are written differently).

speccoder 0.5.2

Fixed

• A Python project's acceptance tests run with its own pytest (the
  check's command, e.g. uv run python -m pytest), not python3 -m pytest, which fails where the system Python has no pytest.
• A failed verdict says what the failing check said first
  (check `acceptance` failed: No module named pytest), at the stop
  and in task show.

speccoder 0.5.1

Fixed

• An attempt the model reports as unfinished stops with
  ATTEMPT-INCOMPLETE and quotes what the model says is missing, instead
  of BUDGET-EXHAUSTED and advice to raise [budget].
• The attempt is told the configured checks' exact commands, so it runs
  them instead of guessing (pytest, python -m pytest) and reporting
  unfinished work because none ran.

Upgrading

• speccoder replay of a run recorded before this release reports a
  divergence (the attempt's prompt changed).

speccoder 0.5.0

Added

• Agents as executors, governed call by call. [executor] agent = "copilot" (or opencode, claude, codex, gemini; any ACP agent
  with kind = "acp" and argv) runs the agent over the Agent Client
  Protocol. speccoder puts it into its ask-first mode and decides every
  permission request itself: everything the agent does on its own asks
  you (you can allow its plain reads for the task), and its file and
  terminal calls go through speccoder's protected paths, masking and
  sandbox. A change it makes without asking is reported as decides
  alone; a missing sign-in stops with EXECUTOR-LOGIN and the agent's
  own login as a step.
• Pick an agent. speccoder run --agent opencode and the app's Run ▾
  menu for one run; doctor --agents and Settings › Executor show each
  agent as installed, signed in and asks first; the live card shows its
  plan and the tally (asked · allowed · waited · refused); a failed task
  offers Try with the next agent.
• Every task finds its doer. A task naming only paths in its
  feature's folder is planning work: it may write design documents there,
  never spec.md, tasks.md, review.json or the constitution, and
  ticks regression_only (planning_work). A [YOU] task, or one that
  stopped BLOCKED, waits for you; the board and Before you run show
  who does what.
• No stop is a dead end. A BLOCKED stop offers I did it,
  Hint & retry and Skip: one key at a terminal (d/h/s),
  buttons in the app, task done / task hint on the command line.
• Run it here. In the app, a fix that is a speccoder command runs or
  opens its view; your own fix (npm install, git add … && git commit)
  runs from a card after a yes, in the OS sandbox with network, only as
  speccoder built it. Sign-ins open a terminal. Every error has Copy
  (masked) and Copy Bug Report; --copy at the CLI.

Changed

• [executor] kind = "claude" and "codex" are gone: write
  agent = "claude" / agent = "codex". kind = "command" stays.
• A task line's leading / (/specs/001-x/plan.md) names the
  repository root.

Fixed

• An acceptance author's command given as one string
  (["cargo test --test acceptance_us1"]) was refused, so new projects'
  tasks ticked regression_only; it is now that command's words.

Upgrading

• speccoder replay of a run recorded before 0.5.0 reports a divergence
  (model calls are recorded with a different key order); ticks and
  attempts in git are unaffected.

speccoder 0.4.0

Added

• Amazon Bedrock as a model provider. [model] provider = "bedrock"
  with a region. Credentials come from AWS — a Bedrock API key
  (AWS_BEARER_TOKEN_BEDROCK) or an aws login/SSO profile — never from
  the file. init --provider bedrock [--region] writes
  global.anthropic.claude-sonnet-5-5; doctor names the source.
• A run's worktree gets its dependencies. Before the baseline, a Node
  or uv project is installed in the worktree with the toolchain's locked,
  offline install, once per lockfile and again when an attempt changes it
  (PREPARE-FAILED when it cannot; [worktree] prepare = "off" skips it).
• A run's worktree gets the git-ignored files its tests need.
  [worktree] include = [".env"] and an existing .worktreeinclude copy
  them from your checkout. Secret-named values in a copied .env, and the
  lines of a copied key file, are masked as [redacted:NAME] in what the
  model, the journal and the screen receive; a commit never holds a copied
  file, and a diff that adds such a value stops the task
  (SECRET-IN-DIFF).

Changed

• Built on agentplane 0.55.0.

Fixed

• A Node project stopped at the baseline in a run's worktree, which has no
  node_modules/; the worktree is now prepared first.
• A uv project's checks ran .venv/bin/python, which a worktree never has;
  init now writes uv run --locked --offline python -m pytest …, and a
  check run through uv run, poetry run or uvx allows exactly itself.
• An empty pytest suite (exit 5) failed the baseline as a broken check; it
  now passes as no tests yet, as nextest's --no-tests=warn does.

speccoder 0.3.0

Added

• Start a project from one line of intent. speccoder new (and New
  Project in the app) plans git init, the stack's own scaffold (cargo,
  uv + pytest, npm + vitest, go), the format's own init (Spec Kit with the
  speccoder preset, or OpenSpec), speccoder init --yes and one commit as
  you. Every command is shown, network steps name the hosts they reach, and
  each runs sandboxed. The run is bound to the plan's digest and resumable;
  --undo moves an untouched folder to the Trash or reverts the commit.
• Plan through: speccoder spec specify "…" --through tasks (and
  issue start N --through tasks) runs the format's own skills in order,
  stopping only for open questions and your approval of the spec.
  feature similar finds a requirement a description duplicates.
• The composer and New Feature. ⌘N suggests the rung — bug fix, quick
  change, feature — from the words, with its reason; the New Feature sheet
  shows the skills that will run, what they write and what they cost.
• A welcome outside a project: New Project, Open Project…, Clone from
  GitHub…, Try the Sample, recent projects. speccoder sample replays a
  recorded run strictly, without a model key.
• A test file the runner never collected fails the verdict
  (TESTS-NOT-COLLECTED), decided from the toolchain's rule (Cargo, pytest,
  vitest/jest, Go), the check's JUnit report, or a listing from the check's
  own command (or its list command). When speccoder cannot tell, the task
  stops, never ticked (TESTS-COLLECTION-UNKNOWN).
• Every regression-only tick says why — no story, no scenarios,
  acceptance off, no usable tests, none collected, not red first, or already
  passing (naming the task that implemented it) — in the trailer, the signed
  statement, status, report, the board, the TUI and ACP. A feature with
  mostly regression-only ticks gets one Little Was Verified item.
• Tests go where the runner finds them. speccoder detects the test
  layout and gives it to the implementer and the test authors; its own
  acceptance tests go to one file per story.
• Plan checks before the run (doctor --feature, Before you run):
  tests named where the runner won't collect them, a story's behaviour in
  Setup or Foundational, tasks that only write tests. speccoder task acceptance-at T --story US writes a story's acceptance tests at that task.
• speccoder preset add: a bundled Spec Kit preset whose tasks template
  plans for evidence, installed by specify preset add. Spec Kit's scripts
  get the installed specify's Python (SPECKIT_PYTHON_EXECUTABLE), which
  has PyYAML; doctor warns when there is none.

Changed

• Built on agentplane 0.54.0.
• One binary. speccoder-verify is now speccoder attest verify --branch (with attest tour and attest badge); the evidence Action
  downloads speccoder. The check and its payload are unchanged; exit codes
  are speccoder's (0, 26 failure, 15 neutral). --require-signed is
  on by default with --branch.
• Watch a project while a run works on it. status, report --approvals, inbox, task show --account and the app read a running
  journal at once instead of waiting. Only a second writer is refused, at
  once. A run killed before closing its journal leaves it for the next
  writer to repair; the listings say so.
• OpenAI's words beside its calls are shown (Responses API), live, in
  waiting calls and in task show --account.
• speccoder eval stops on a refused account by the provider's error
  code, not by words in the message. Not yet recognised: Anthropic's
  "credit balance is too low" (invalid_request_error).

Fixed

• A call that could not be asked about (--no-prompt) said only "not a
  configured check or allowed command"; it now names the command and how to
  allow it.
• review approve --run and run on a feature without tasks said "feature
  done"; now NO-TASKS (exit 16) names speccoder spec tasks.
• On macOS, mktemp failed inside the sandbox, so Spec Kit's scripts could
  fail.
• A planning skill that stopped incomplete gave no reason; it now quotes the
  skill.
• Flags after a description were read as part of it; a description starting
  with - goes after --.
• Reading a project (task show --account, a replay, doctor) no longer
  creates a journal or the state directory.
• The test suite no longer writes to your real state directory.

Upgrading

• A 0.2.0 journal still opens and its runs still list and show. A decision
  recorded under 0.2.0 does not read in this build: the attempt that waited
  on it fails once, saying so, and the next speccoder run asks again.
  Ticks and attempt refs in git are untouched.
• Upgrade every speccoder that works on a project together: a 0.2.0
  process holds the journal alone.

speccoder 0.2.0

Added

• Bug fixes as their own lane. speccoder bug "<what happens>" --expected "<what should>" records the bug where the format keeps them (Spec Kit's
  bug extension, else a standing bug unit), writes one reproduction per
  expected behaviour that must fail by an assertion, asks you once to
  confirm it, and ticks the fix only when the reproductions pass, the
  unchanged behaviour still passes, the held-out variants pass and the
  revert is rejected. cannot_reproduce and already_satisfied never tick.
  --test brings your own failing test, --candidate PATCH|pr:N judges
  someone else's fix, --issue N closes the issue when the fix is verified.
• Review the oracle. speccoder review tests <REQ> writes a
  requirement's acceptance tests during review: two independent authors,
  disagreements shown as ambiguities, red-checked in the sandbox, a held-out
  set sealed. speccoder review approve --run ask|autopilot freezes them and
  starts the run, which reuses them without another author call.
• Anchored comments. Comment on spec passages, tests and findings, send
  them as one revision; Undo Revision restores the documents byte for byte.
• speccoder-verify and the speccoder / evidence check. A small
  binary — no model, no network — checks every ticked task of a branch
  against its signed evidence with the maintainer's trust. A GitHub Action
  publishes the result as a check run on the pull request.
• The evidence tour of each task — requirement, frozen tests, red, diff,
  green, strength — in the check, the pull request, the app and
  speccoder-verify tour. The task-verification predicate is documented,
  with a JSON Schema.
• speccoder actions [--task T] [--json]: what can be done now, the list
  every interface offers.
• An asking budget: after [approvals] budget prompts (default 12) further
  calls wait together in one Needs You item (speccoder approvals batch).
• The model's reasoning summary (Anthropic, OpenAI Responses), shown as its
  account, never as evidence.
• task restore and task drop for changes a run set aside.
• Setup gaps a run would trip on (an uncommitted lockfile, unignored build
  output) stop it with a one-click fix (SETUP-INCOMPLETE, speccoder fix).
• The app shows runs started in a terminal, and untracked and ignored files
  in its Files view.
• cargo binstall speccoder.

Changed

• Approvals cover exactly what you saw — the file a write reaches, a
  command's exact argv with its directory and program, an MCP tool's
  definition — are checked again when the call runs, and end with their
  run.
• [commands] allow and [sandbox] network_allow match exact commands.
  Prefixes go in allow_prefix / network_allow_prefix and are saved only
  after a preview of what they would allow (speccoder config coverage).
  Always Allow adds exactly one command. An approvals.json from 0.1.0 is
  refused; speccoder doctor says so.
• A write that removes a test is refused before it happens.
• speccoder pr also pushes the held-out tests of the closed tasks and adds
  each task's tour to the description.
• Decisions record the interface they were made in (terminal, web,
  acp).
• A feature's branch takes setup changes (configuration, .gitignore,
  lockfiles) from the checkout.
• error[CHECK-BROKEN] quotes the check's own error.
• agentplane 0.53.0.

Fixed

• Needs You, inbox and status said nothing was waiting when another
  speccoder held the journal or the decisions could not be read; they now
  say the calls are unknown, and why.
• An OpenSpec project without a change was offered no first step.
• With CARGO_TARGET_DIR set, every Rust check failed in the sandbox (it may
  not write outside the worktree); checks build in the worktree instead.
• The app refused its own concurrent journal reads.
• attest verify reported a missing cosign as an untrusted signer.

Security

• On a case-insensitive volume (macOS) another letter case reached a
  protected file (Speccoder.toml, .CURSOR/…); protected paths now match
  in any case.
• A write through a symlinked directory could reach a protected file;
  protection is judged on the file a write reaches.
• An approval could reach more than was shown: an Always Allow that widened
  to a prefix, a path grant a symlink carried to another file, an MCP grant
  that outlived a changed tool definition, a task grant that seeded later
  runs.
• Approve and Reject in the app, the dashboard and Needs You decided
  whichever call waited for the task; they now name the call and its
  digest, and a changed call is refused (CALL-CHANGED).
• A command whose arguments the model copied from content it read can be
  approved once only.
• Claude Code as the executor got each allowed command as a prefix rule
  (Bash(cargo test:*)); it gets the exact command (Bash(cargo test)).

speccoder 0.1.0

The first release. speccoder implements GitHub Spec Kit features and
OpenSpec changes task by task, and ticks a task only on evidence the model
cannot produce or edit.

Added

Judged by evidence

• Per task: baseline checks, spec-derived acceptance tests that must fail
  first (red) and pass after (green), the project's checks again, and a
  diff audit for weakened tests, skips and config edits. Verdicts:
  verified, regression_only or failed; a task the model cannot do
  honestly stops as blocked.
• Acceptance strength: plausible wrong versions of the change must fail
  the tests; a test that rejects what the spec allows goes to a person.
  Optional held-out tests, mutation testing and a gaming monitor.
• Every tick is a commit with its evidence in trailers and an in-toto
  attestation in git notes (speccoder attest verify).
• When the spec changes, ticks proven against the earlier wording say so,
  word by word; run --converge re-checks the code against the spec as it
  is and adds tasks for what no longer holds.

Safe by construction

• Work on speccoder/<feature> in its own worktree, or in place.
• An OS sandbox for every command: Seatbelt on macOS, Landlock and seccomp
  on Linux — writes only to the worktree, no credentials, no network
  beyond loopback, no Unix sockets outside the workspace.
• A policy classifies every tool call before it runs: allowed, sent to a
  person, or denied. Approvals for once, the task or the run, inline, in
  an editor (ACP), in the app or later — each recorded with who decided.
• Every model and tool call journaled by agentplane: a killed run resumes
  without repeating finished calls; any recorded run replays offline.

The person's work

• Needs You (speccoder inbox): every decision waiting in every feature,
  the most urgent first, each settled in place.
• Spec review (speccoder review): approve each requirement's wording or
  request changes; answer open questions in place; reword a requirement;
  see what changed and the spec's history. [review] required makes a
  run wait for it.
• Quick changes (speccoder do "…"): a story, an approved requirement and
  a task in a standing feature, committed and run like any task.
• The model's account: what the model says beside its calls, live and
  afterwards (task show --account), always shown as a claim.

Interfaces

• A terminal dashboard for run and bare speccoder; --plain, screen
  reader mode, ASCII and colour themes.
• The app (speccoder ui, --window for a native window):
  Needs You, Overview, Board, Code, Spec & Plan, Report and
  Settings; a menu bar, a status bar, context menus, a command palette and
  a setup assistant. Local only, on 127.0.0.1 with a per-start token.
• Editors over ACP; MCP servers as tools, pinned by digest.
• --json and --events on every command with published JSON Schemas;
  every failure carries a stable code and the command that fixes it.

Spec formats and planning

• Spec Kit 0.13 to 1.2: its planning skills hosted with speccoder's model
  (speccoder spec specify|plan|tasks|…), extension hooks, and a step
  type for Spec Kit workflows (speccoder workflow install).
• OpenSpec 1.0 and newer: changes, spec deltas, openspec archive
  preflight, speccoder spec propose|archive.

Models and agents

• Anthropic, OpenAI and any OpenAI-compatible server (Ollama, vLLM, …);
  keys from the environment only. Claude Code, Codex or any command as
  the agent making the attempts, judged the same way.
• Parallel [P] tasks in lanes, best-of-n attempts, retries with the
  verifier's facts, token and cost reporting.

GitHub, without gh

• speccoder auth login through speccoder's GitHub App (a short code in
  the browser; keychain storage); issue start 42 specifies a feature
  from an issue and can comment on it; pr opens a draft whose
  description is the evidence.

Distribution

• Release archives for macOS (Apple silicon) and Linux (x86_64, aarch64)
  with SHA-256 checksums and build-provenance attestations; a shell installer
  (--verify-attestation); a Homebrew tap; the crates on crates.io.
• Documentation at https://hupe1980.github.io/speccoder/.