Skip to content

0.10.0 — skill libraries, GitHub review loop, harness & skill-bank tooling

Latest

Choose a tag to compare

@Birfy Birfy released this 13 Nov 08:26
f6a6e3c

Added

  • agentdescent github publish|status: publish one accepted candidate as a
    proposal branch, commit, pull request and optional check run, and read a PR's
    state and checks back. Publishing is a plan by default (--apply writes), the
    token comes only from an environment variable, the base branch is never written,
    and an existing proposal PR is updated rather than duplicated. The PR body
    carries the run/artifact ids and evidence hash; there is no auto-approval,
    auto-merge or branch-protection change.

  • agentdescent github review: read a PR's reviews, inline comments and check
    runs into one packet, extract classified review constraints (test, security,
    api, docs, style, behavior), and report merge readiness with blockers. It is
    read-only, never approves or merges, and requires an explicit requires_human_click
    step; CHANGES_REQUESTED and its inline comments are blocking, plain notes are
    not turned into tasks, resolved ids can be supplied because REST does not
    expose thread resolution, and only a reviewer's decisive review counts (a later plain
    comment never lifts a block), so a change request that reviewer later
    approved or had dismissed no longer blocks.

  • agentdescent skills report|audit|search: a durable skill bank with an outcome
    ledger (uses/successes/failures), a bank-relative lift audit that
    quarantines underperforming skills (read-only unless --apply), active-only
    retrieval through the artifact graph, and explicit retire/activate lifecycle
    states. The store is an atomically written JSON snapshot; no model or network
    call is made.

  • agentdescent hgm next|report: HGM's missing evaluation policy and
    expand-vs-evaluate scheduler over a JSON tree. The planner applies the
    n_task_evals ** alpha budget split, Thompson sampling on clade evidence for
    expansion, own evidence for evaluation, optional cool-down, and unsubmitted-task
    selection; it stops at the budget or when every task is spent. Utility measures
    are validated as [0, 1]; no model or network call is made.

  • agentdescent harness manifest|check and HarnessManifest: a file-level
    component contract for harness evolution. Components name path globs, an RRSI/SHE
    responsibility, editable/frozen state and per-component file/line budgets; the
    manifest also has global budgets, allowed/forbidden paths and a FileTree kwargs
    projection. check fails closed on frozen, forbidden, unowned or over-budget
    edits; no model or network call is made.

  • agentdescent compile-skill and compile_skill: turn a skill document into a
    single tested function through a caller-supplied text -> source compiler,
    behind the framework AST gate, with repeatable tests and optional
    reference/case equivalence. Any gate, test or equivalence failure returns
    status: "text" with a reason and leaves the skill document as the fallback; no
    model or network call is made by the module.

  • Documented the contributor decision tree for Algorithm card vs MethodPolicy vs full examples/ port vs aggregator_factory (docs/algorithm-decision-tree.md), cross-linked from authoring / port-fidelity / aggregator-factory. plan now warns when acceptance/promotion (as well as conflict/fusion) are installed on a factory path and will be ignored — engine composition without silent drop remains #807.

  • kind=skill_library is a first-class EvolveSpec / CLI / MCP kind: FileTree +
    tree_runner(layout="skill_library"), init --kind skill_library seeds ≥2
    */SKILL.md, and plan echoes the resolved layout. See
    examples.skill_library_evolution (#820).

  • agentdescent apply --install-to PATH (and MCP install_to) installs an evolved
    skill_library into a host skills root such as ~/.claude/skills with the same
    write_to backup semantics and without pruning unrelated host skills (#823).

  • agentdescent demo --algorithm runs offline for gepa, dgm, ace, skillopt and
    rrsi on a text fixture (archive_selection stays on the skill_dir demo);
    evoskill is refused with an explicit reason (#845).

  • agentdescent demo --kind skill_library runs the offline multi-file
    */SKILL.md library demo from an install (no source checkout, no key).
    The default demo smoke note and README / quickstart-skill / usage now
    point at that command; python -m examples.skill_library_evolution stays
    as the source-checkout CLI (--dry-run, --agent, --install-to) (#878).

  • status, watch, show, apply, cancel, resume and diagnose accept
    latest (or last) in place of a run id and use the newest run in the store,
    so agentdescent show latest works straight after a demo; an empty store
    says so instead of no run 'latest' (#892).

  • MCP Path A now has a demo tool (algorithm, optional kind) that returns
    the same payload as agentdescent demo --json, so a host can run the zero-key
    smoke without a provider key. The shared skill tells hosts to try demo (or
    agentdescent demo --json) before doctor → plan → start (#903).

  • The read-only run store streams a run's status and new rounds over Server-Sent
    Events at GET /api/runs/<id>/events, with Last-Event-ID resume, so a host
    panel or platform can follow a run without polling (#922).

  • agentdescent demo --algorithm evoskill (and demo --kind skill_library --algorithm evoskill) now runs the evoskill card offline, with no API key: the demo reflector speaks the Proposer/Generator protocol and induces a */SKILL.md with frontmatter, held-out reward rises, and outcomes include evoskill-select. Documentation no longer says evoskill is refused in the offline demo (#934).

  • RunManifest.replay_eligible / replay_notes and an optional replay record
    (agentdescent.replay_fidelity.ReplayFidelity) that captures what a replay needs
    and the manifest's input hashes do not cover: RNG seed/state digests at
    interaction boundaries, an environment fingerprint (cwd, interpreter, tool
    versions, and the names of the allowlisted environment variables present --
    PYTHONHASHSEED,
    PYTHONPATH, LANG, LC_ALL, TZ, OMP_NUM_THREADS), and hidden-state
    digests. No value material is stored (not even hashed), a caller cannot widen the
    allowlist, and a credential-shaped name is refused even when allowlisted. Trajectory carries the same record (with replay_eligible /
    replay_notes properties), ReplayLab(manifest=...) refuses a counterfactual
    with ReplayNotEligible naming what was not reconstructed, and replay_check()
    refuses to compare two runs when either side was not captured (#939).

  • A run, a commit and a decision can now name the decision procedure that produced
    them, as a content hash over the Policies bundle, the decision-relevant
    AggregatorConfig fields and the verifier's knobs:
    policy_digest on DecisionRecord, MergeContext, RunManifest (build_manifest)
    and EvolutionResult, plus a policy-digest: trailer on the ledger's commit
    message. A decision taken under a different policy than the previous one for the
    same artifact links to it (parent_id + evidence["policy_change"]), and
    agentdescent.policies.policy_digest() computes the version for a hand-built
    bundle, and StableDistanceAcceptance / AdaptiveTrustRegion now expose their
    own knobs to it. A decision record that does not state a policy keeps
    policy_digest empty rather than claiming the shipped default, and
    run_guarded() records the digest its result carried. Additive: empty by
    default, unchanged commit messages and unchanged manifest_id (#943).

  • A promotion decision now sees what reviewing the change costs a human: the new
    ReviewCost records the diff's files, lines, blast radius, coarse risk tier and
    evidence count and estimates review minutes from documented constants, and
    PromotionRules gains the opt-in max_review_cost / max_review_minutes limits
    (both default off, so with them unset the decision and its hashes are unchanged)
    that defer a disproportionate change and recommend batching it. The decision's
    evidence reports attention_per_accepted_gain when a cost and a positive eval
    delta are present (#950).

Changed

  • Hard cards vs ports contract: Path A names wheel Algorithm cards only;
    examples/ research ports stay source-only CLI modules (same slug ≠ same
    fidelity). Decision tree, README, plugins, SKILL.md, and port-fidelity updated;
    future MethodPolicy card candidates listed as scoped follow-ons (#384).

  • algorithms list / show, MCP list_algorithms / describe_algorithm, and plan's algorithm echo now label every registry entry as an Algorithm card (surface=card), filter-friendly by host kind, and — for ACE / GEPA / DGM / SkillOpt / EvoSkill — name the sibling examples/<algo>/ research port with an honest fidelity_note so same name is never read as same fidelity.

  • The homepage interactive demo now replays a session the way the shipped skill runs one: each host's real slash command (or plain request for DeepSeek Harness), default skill folder and MCP tool naming; doctor, spec, plan with the real call estimate, a start approval, status polls with a round/phase/held-out progress panel, show diff=true with a unified diff, and a separately approved apply with the real backup location. Still a browser-only sample; no model calls.

  • The README is reorganised around install, a 60-second offline quickstart,
    calling any named algorithm, adding one, and a host table (Claude Code, Codex,
    OpenCode, DeepSeek Harness, Grok Bot). docs/testing-the-plugins.md is merged
    into docs/testing-guide.md, and stale port counts, version references and
    duplicated install text across the docs are fixed (#797).

  • The MCP server now exposes Path A by default (doctor → plan → start → status/show → apply, plus cancel/resume, diagnose, and algorithm list/describe/spec). Audit, release, platform, experience and the other satellites stay callable with the same names when agentdescent mcp --advanced or AGENTDESCENT_MCP_ADVANCED=1 is set. Docs: docs/mcp-path-a.md.

  • evolve / async_evolve accept typed BudgetSettings, RuntimeSettings and ObservabilitySettings bags; flat kwargs stay as shortcuts. Docs and EVOLVE_PARAM_ALLOWLIST freeze the flat surface so new knobs go in Policies or a bag (#804).

  • aggregator_factory= now forwards merge-side Policies into the factory and
    applies them on aggregators that expose the slots; leaving them unused raises
    unless accept_orphaned_merge_policies=True. GEPA and SkillOpt factories honour
    acceptance / fusion (#807).

  • ProposalPolicy may return multiple strings: evolve / async_evolve ingest
    each as its own evidence card, pass ctx.k from policy.k, and no longer raise
    ProposalContractError on len > 1. Self-verify, when enabled, runs per
    proposal (#808).

  • evolve(asynchronous=True) is a deprecated bridge to async_evolve: it now
    requires explicit max_seconds and max_rollouts (no silent 20s default, no
    rounds reinterpretation) and emits FutureWarning. Policies wiring is one
    shared _POLICY_DRIVER_SUPPORT table for both drivers (#809).

  • Flagship skill demo is now examples/skill_library_evolution.py: a multi-file Agent Skill library (layout="skill_library", offline agent must read skills off disk, --dry-run / --install-to). examples/skill_evolution.py and README / docs/quickstart-skill.md are retitled as instruction / playbook evolution so SingleSlot / AppendRules are not called "a skill" without qualification (#818).

  • MethodPolicy's keyed string bank is renamed KeyedSkillBank (alias SkillLibrary
    kept). Voyager / SkillWeaver docs and imports no longer call it an Agent Skill
    library; disk libraries point at skill_library_evolution / EvoSkill (#821).

  • Decision tree and aggregator-factory prose match closed #807 (compose/raise, no
    stale “until”); quickstart-skill titles say instruction / one skill / library;
    README lists policy_slot and points at #820 for skill_library; GEPA docs /
    print paths say prompt/instruction, not skill folder (#822).

  • algorithm: evoskill on kind=skill_library selects SkillLibraryTree LIBRARY
    layout (skills/*/SKILL.md); skill_dir stays SINGLE_SKILL. plan notes name
    the layout. Card kinds include skill_library (#824).

  • ACE card notes and summary state the ## general fallback and that admit means
    "cleared Jaccard dedup"; FiNER templates are labeled example-only; extra_calls
    no longer says none while describing a reflection call (#842).

  • Docs state that Path A reported held-out reward is the select-set score (optimistic); a locked test split remains #843.

  • evolve(test_frac=...) can carve a test partition out of the held-out tail:
    accept/select keep reading the select split, the test split is scored once at
    the end and reported as test_reward, so the printed reward is no longer the
    score of the set that chose the winner. With no split, evolve, show and
    diagnose now say plainly that the reported reward is the select set, not an
    unseen test set (#843).

  • A text-only Algorithm card asked to run on a skill_library (or any other tree
    kind) now says so in the kind-mismatch error, before the list of cards that do
    apply: 'gepa' is text-only: it reads a single text file, not a skill_library target. The list of applicable cards and the one-file --kind text route are
    unchanged (#846).

  • Algorithm decision tree notes which cards are text-only vs skill_library-capable (#846).

  • agentdescent demo labels itself a smoke trajectory and points at
    examples.skill_library_evolution for multi-round library edits; that offline
    path now edits a */SKILL.md body (STATUS draft→ready) and
    init --algorithm evoskill defaults batch_size to 1 (#847).

  • init --kind skill_library tips name the archive_selection Path A card
    (and algorithms list --kind skill_library); quickstart-skill, decision-tree,
    and algorithm-authoring kinds stay aligned (#864, Refs #846).

  • An unknown algorithm name that is a near-miss typo of a registered card now ends
    with "did you mean 'gepa'?" (algorithms show / run, specs, MCP), instead of
    only listing every card (#874).

  • agentdescent plan and the plan step of algorithms run print a short human
    summary by default -- task and round counts, the agent-call upper bound, the
    agent, the small-held-out warning evolve raises, and the next: command --
    instead of the whole plan document. --json is unchanged (#919).

  • TrustEvidence can now carry a RunHealth record (attempted,
    failed_tasks, threshold): failed tasks are dropped from the paired scores
    instead of being folded in as zeroes, so the reported gain is the mean over the
    tasks that answered, and a run whose failure rate is above the threshold fails
    its eval gate, so promote() cannot return promote. A bundle built without
    health data keeps today's keys, values and hash (#993).

Fixed

  • The skill-library seed (init <target> --kind skill_library and
    algorithms run <name> --kind skill_library) now refuses an unwritable target
    as one error: cannot write <path>: <reason> line with exit 2, or
    {"ok": false, "error": ...} under --json, instead of a traceback. Both
    destinations are preflighted before anything is written, so a refused --out
    leaves no seeded */SKILL.md behind; a writable target still seeds its
    starters and the #868 one-starter note is unchanged (#1003).

  • An expired or unprovisioned subscription refused by the provider -- Ark
    Coding's HTTP 400 InvalidSubscription, OpenAI's account_deactivated,
    Anthropic's billing_error -- now fails the first call as SubscriptionInactive
    instead of being retried as a transient RuntimeError. The run stops within
    seconds instead of after three attempts per call, the error line names the
    subscription as the cause, diagnose repeats that next step instead of
    recommending doctor, and resume refuses it without --force (#1006).

  • --budget / budget_usd is only checked between rounds: plan, evolve, and
    MCP start now warn when the budget is below the seed+one-round call upper
    bound (so the first round alone can overshoot), and the plugin quickstart /
    plugins docs match that semantics (#764).

  • ReflectiveFusion now ranks a synthesised union against its parent singles before the acceptance gate, so a merge that loses to a stronger parent cannot commit just because it still beats the seed (trust_union=True keeps the older cheap unranked path).

  • ReflectiveFusion's MERGE_PROMPT now receives compact parent cheap-score and rationale metadata, and instructs the model to pick one coherent parent for incompatible taxonomies/enums instead of OR-concatenating them (#767).

  • plan / evolve / MCP start now hard-refuse a bare isolated opencode worker
    when a gateway key (e.g. ARK_API_KEY) is set, instead of soft-warning and then
    burning calls on OpenCode's unactivated Volcengine default. Init/algorithms run
    warnings say filling reflect alone is not enough — pass --model / --base-url
    (or write the Coding recipe). Wired --model/--base-url specs stay green (#790).

  • Missing-CLI plan/algorithms-run errors for directory kinds no longer suggest
    openai_compatible (which those kinds refuse); they point at WorkspaceAgent
    CLIs instead. kind=text / reflect remedies are unchanged (#791).

  • agentdescent algorithms list --kind policy_slot now prints the same explanation
    MCP list_algorithms already returns (that kind evolves a policy slot and plan
    refuses an algorithm), instead of a bare "no algorithms for kind" line. --json
    keeps the array shape and puts the note on stderr (#799).

  • agentdescent algorithms show / run (and MCP describe_algorithm /
    algorithm_spec) now tip that a documented examples/ research port
    (reflexion, textgrad, self-refine, …) is not a named Algorithm card —
    pointing at docs/algo-*.md and the checkout python -m examples.<slug>…
    path — instead of only listing the seven registered cards (#801).

  • PopulationAggregator now archives per-task held-out scores and fills
    Candidate.per_task, so Policies(selection=ParetoFrontier(...)) can run
    GEPA-class parent selection without aggregator_factory= (#806).

  • doctor keeps tipping Grok Bot operators to save the scaffold skill as a user skill when user_skill is missing (and appends the same follow-up on a stale/partial scaffold problem). The Bot loads the saved user skill, not the scaffold alone (#825).

  • init / algorithms run writing an Ark (*.volces.com) openai_compatible block now set "stream": true by default so reasoning models like deepseek-v4-flash do not hit the 120s non-streaming propose timeout; plan warns when a hand-edited Ark block omits it, and the quickstart / plugins Ark examples match (#832).

  • The offline agentdescent demo sample output in the plugin quickstart and
    testing guide now shows the plan line the CLI actually prints (up to 68 agent calls, not 64) (#833).

  • agentdescent show / show --json (and MCP show) now surface
    fusion_stats / fusion_summary from reflective-merge ranking
    (contested, unranked, losses) so Path A can see merge health after #766
    without the ledger Python API (#834).

  • kind=text now uses SeededAppendRules so a one-sentence reflector proposal
    appends under # Learned rules instead of replacing the whole seed prompt
    via SingleSlot (#837). GEPA keeps SingleSlot with a full-rewrite propose
    template.

  • evoskill on kind=skill_library writes top-level <slug>/SKILL.md (with
    Agent Skills frontmatter) instead of nested skills/<slug>/SKILL.md, so apply
    into a host skills root stays loadable and edits reuse the seed path (#838).

  • GEPA no longer stores unused conflict/promotion/screen Policies as
    attributes, so #807's orphan check fails closed instead of reporting them
    as active (#839).

  • SkillOpt LRScheduler clamps the anneal fraction to [0, 1] so the edit
    budget stays at min_lr after total_steps instead of climbing back (#840).

  • RRSI card defaults use T=8, a small noise_band, and non-zero betas so a
    short evolve run still exercises the anneal and cost gates (#841).

  • DGM increments children only for archive admits; unused evoskill val_every
    removed; GEPA SelectionContext.head tracks the last selected parent (#844).

  • algorithms run error paths dedupe identical init/plan tips, and --help
    names Path A vs algorithms (#848).

  • agentdescent demo --json now emits a parseable result object (ok, run_id,
    rewards, learned, plan, next) instead of ignoring the flag and printing the
    human narrative (#862).

  • agentdescent init <path> and agentdescent algorithms run NAME <path> now
    guess kind=skill_library for a directory with two or more */SKILL.md and
    no root SKILL.md, instead of falling through to skill_dir (#866).

  • init --kind skill_library / algorithms run on a library path now name only
    the starter */SKILL.md folders actually written (e.g. example-alpha when the
    target already had one skill), instead of always claiming example-alpha / example-beta (#868).

  • When a GSM8K research port cannot fetch the dataset, the error now ends with
    what to do next: pip install pyarrow (when that is why the parquet path
    failed), set HF_ENDPOINT to a reachable mirror, or set
    AGENTDESCENT_GSM8K_SAMPLE=1 for the bundled offline sample (#870).

  • The missing-reflector-key warning from plan / algorithms run now names how
    long the run really lasts: "fails after one round" on a --rounds 1 quick
    trial instead of always "two rounds" (#872).

  • Unknown algorithm parameter names now fold case and -/_ (batch-size →
    batch_size) and a near-miss typo ends with "did you mean 'batch_size'?",
    matching how algorithm names already resolve (#876).

  • When a kind=text card or caller sets strategy, plan no longer also
    claims SeededAppendRules. gepa / ace / dgm / skillopt drop the
    generic note; rrsi and a bare text spec still show it (#879).

  • --kind and spec kind now fold case and -/_ (skill-library →
    skill_library) and a near-miss typo ends with "did you mean 'skill_library'?",
    matching how algorithm and parameter names already resolve (#882).

  • --agent now folds case and -/_ (claude-code → claude_code,
    openai-compatible → openai_compatible), matching how algorithm names,
    parameter names, and --kind already resolve (#884).

  • An unknown CLI verb no longer lists the internal run worker among
    choose from choices, and a near-miss (algoritms, compile_skill)
    gets did you mean …? (#886).

  • Algorithm parameter values that declare choices now fold case and -/_
    (sigmoid-novelty → sigmoid_novelty, KEEP_ALL → keep_all) and a
    near-miss typo ends with "did you mean 'novelty'?", matching how
    parameter names already resolve (#888).

  • uninstall no longer tells grok-bot, dsh, or OpenCode users to remove files
    via a host marketplace those hosts do not have. The marketplace aside stays
    only for Claude Code and Codex, which can be installed that way (#890).

  • init --algorithm writes the registered algorithm name into the spec, so
    --algorithm GEPA or archive-selection is saved as gepa /
    archive_selection (matching algorithms list and algorithms run) instead
    of the alias as typed (#894).

  • show_payload / status_payload / apply_payload and the MCP show / status / apply / diagnose / cancel / resume tools accept latest (or last) the same way the CLI does, so a host that asks to show the newest run no longer gets no run 'latest' (#896).

  • agentdescent diagnose now names the runnable next command with the resolved run id for solved (show <id>, apply <id> --dry-run), cancelled (resume <id>), running (watch <id>) and partial-improvement (show <id> --json), instead of prose such as "review and apply the artifact" (#898).

  • agentdescent resume (CLI and the MCP resume tool) now refuses a run that is already done [target_reward] and points at show <id> / apply <id> --dry-run, instead of relaunching it, spending calls again and overwriting its recorded initial reward and outcomes with the no-op rerun. Runs that stopped on a bound, were cancelled or failed still resume (#900).

  • agentdescent demo --algorithm NAME now ends with the one command that runs that algorithm on your own data (agentdescent algorithms run NAME <your-prompt.txt|your-skill-dir> <your-cases.jsonl> --model <model> ..., then agentdescent evolve agentdescent.NAME.spec.json), in both the text footer and --json next, instead of a generic init that dropped the algorithm. demo without --algorithm is unchanged (#902).

  • agentdescent demo no longer leaves an agentdescent-demo-* directory in the
    temporary path after every bare run: the default root is removed on the success
    and the error path, while --dir keeps the directory it names and --json
    still reports a dir that exists. The quickstart line now says which case keeps
    the files (#906).

  • A run store that cannot be created (--store, AGENTDESCENT_HOME or the default
    ~/.agentdescent/runs on a read-only, missing, unwritable or file path) is now a
    RunStoreError: the CLI exits 2 with one error: cannot create run store ...
    line, or {"ok": false, "error": ...} under --json, instead of a raw
    FileNotFoundError / FileExistsError / PermissionError traceback (#907).

  • The install FAQ's port taxonomy now names all four catalogue groups, Genesis
    included, so the dry-run split adds up to the 23 ports the same paragraph
    quotes instead of the 8 + 14 = 22 it printed (#908).

  • agentdescent serve --host <non-loopback> now refuses to start, with one
    error:-style line and exit 2, unless the operator passes --allow-remote: the
    run panel has no authentication and lists every run in the store, which a single
    --host 0.0.0.0 used to publish to the network. Loopback hosts are unchanged,
    and the opt-in prints the same no-authentication warning agentdescent mcp
    does (#911).

  • agentdescent github publish now reads the repository's default_branch and
    refuses a proposal that names it, even when the proposal's own base is another
    branch. An existing proposal branch is only fast-forwarded: a rewritten history
    raises one GitHubError and nothing is written, and no ref update is ever sent
    with force: true (#913).

  • The review gate now counts a Verdict: PASS written above its Tip SHA: line,
    above a bare Tip: <sha> line, or carrying (tip <sha>) inline, so a PASS at
    the right tip is no longer read as "says nothing at this tip"; What I ran (...)
    counts as the Ran: list. A PASS with no Ran:/Not run: is still counted but
    marked weak: no Ran:/Not run: (#915).

  • doctor now flags an MCP launcher whose first argument is an absolute path that is no longer an executable file: the host reads mcp broken / state: partial (exit 1) instead of a silently green mcp ok. The host line names the dead path, and the problems line tells you to re-run agentdescent install <host>. A relative or bare command such as python3 is not flagged; Grok Bot stays mcp n/a (#918).

  • agentdescent resume no longer relaunches a run whose last failure was a spent
    provider quota: a failed run whose error starts with QuotaExhausted is
    refused with exit 2 and the provider's own reset time in the message, plus
    --force to retry anyway, and the status file is left untouched (#926). Only
    that case is gated -- a cancelled, bounded or ordinary provider failure still
    resumes. The MCP resume tool takes force=true for the same opt-out.

  • agentdescent demo --algorithm rrsi printed what it learned as the internal
    state mapping (state -> {'ra4bfacbc48': 'COLUMN: amount'}) in the human footer
    and in the --json learned field; both now name the rule text
    (rules -> 'COLUMN: amount'), join several rules with | and count any
    beyond the first two, so no internal id or Python repr reaches the reader and
    the line matches what show renders (#927).

  • The review gate no longer treats a later Ran: / What I ran: block as
    strong just because an earlier block in the same body had a Not run: line;
    hasNotRun now resets when a later list starts, the same way ran already
    does (#931).

  • agentdescent harness check and the FileTree it configures now share one
    matcher (filetree.match_any), so a manifest cannot mean two different things
    on the two surfaces. Every matching component applies -- read-only wins over
    editable and a frozen catch-all can no longer be hidden behind an earlier one --
    budgets are charged to each matching component, and * no longer crosses /
    (** means any depth) (#932).

  • github review no longer reports ready: false because a reviewer's approval
    body or a lifted (dismissed) review contains an action word such as add,
    required or cannot: the action-word promotion is now keyed on the review
    state and applies to comments only, and a review body that is itself a
    Verdict: PASS report is never blocking. Those rows are still listed as
    suggestions, and CHANGES_REQUESTED (with its inline comments) still blocks
    (#933).

  • agentdescent apply --install-to (and MCP apply(install_to=...)) now resolves known host names and aliases (claude-code/claude, codex, opencode/open-code, dsh/deepseek-harness) to their host skills root using the same path helpers and --home / host env rules as install and doctor when no matching path exists in the working directory, while grok-bot/grok refuses writes, exits 2, and prints manual user-skill save paths (#935). agentdescent apply --json also emits structured {"ok": false, "error": ...} payloads on errors like other CLI verbs.

  • build_trust_evidence(..., artifact_changed=False) now reports gain == 0 and
    noise_only is True for an artifact that is byte-identical before and after,
    and such a bundle can no longer return promote: the zeroed eval delta is not a
    pass, so evaluation noise is not repriced as learning (#937).

  • The review skill now takes the merge lane the diff's risk tier gives it: a T0/T1
    pull request is still auto-merged pinned to the reviewed tip, while a T2/T3 one
    gets awaiting-approval and waits for the maintainer's click instead of being
    auto-merged by an agent (#961).

  • append_provenance keeps its records in the ledger's .git/ directory instead
    of the artifact work tree, so git add -A can no longer sweep provenance files
    (and their .lock) into an artifact commit; an artifact_id containing a path
    separator, .. or an absolute prefix is rejected instead of escaping the
    provenance root (#966).

  • MCP apply validates resolved destinations, changed tree entries and backup
    paths against the operator's write allowlist before writing. The shared path
    guard also resolves directory links, including Windows junctions, so an
    allowed path spelling cannot redirect a write outside the configured roots (#969).

  • evolve() cancels workers still waiting for a concurrency slot when their round
    times out, so they cannot start new rollouts after the round or run has ended.
    Calls already in flight may still finish (#970).

  • Async runs now record workers that fail during initialization, task selection
    or completion bookkeeping, and stop with an error when every worker has
    retired. Caller contract errors propagate to the calling thread; rollout
    backend retries and healthy workers keep their existing behavior (#971).

  • EvalOps.run() now uses the default regression gate to set passed, so a
    single-task result no longer reports a pass that EvalOps.gate() rejects.
    Explicit gates that disable significance checks remain available.

  • Use UTF-8 for CLI agent, code/plugin runner and command grader pipes regardless of the parent locale, preserving Unicode answers and JSON grading inputs. Invalid UTF-8 output raises an explicit adapter error instead of corrupting answers or scores (#973).

  • agentdescent serve SSE round events (GET /api/runs/<id>/events) now emit each row's 0-based append ordinal in rounds.jsonl as the event id: instead of row["round"], keeping round in data and ensuring resumed runs and Last-Event-ID reconnects never skip rows (#976).

  • apply refuses a symlink planted in the destination (--to) instead of
    following it: the link itself, or any existing component under --to, is named
    in the error, so a link at skills/SKILL.md can no longer redirect the write -
    or the .bak-N copy taken before it - to a file outside --to. Containment is
    checked on the leaf as well as its parent, and the destination rule now lives in
    filetree.dest_problem beside the tree writer's boundary. The directory
    backup slots scan bak-N with lexists too, so a dangling link there is
    skipped instead of crashing the backup copy (#977).

  • merge-lane reads review-policy.json and risk-rules.json from the PR's base
    commit instead of the clone it runs in, so a branch that edits either file can
    no longer lend itself a PASS or a lower tier; .github/** and
    .claude/skills/** are T3 by rule, and a base that cannot serve either file
    prints no lane at all (#978).

  • A refused artifact write — init --out, algorithms run --out, or
    apply --to for a text file or a tree — now prints one
    error: cannot write <path>: <reason> line and exits 2, or
    {"ok": false, "error": ...} under --json, instead of raising the raw
    OSError as a traceback. The run store's cannot create run store ...
    contract (#907) is unchanged (#990).

  • The last artifact writes outside the run store now report a refused write as
    one error: cannot write <path>: <reason> line with exit 2 instead of a
    traceback: init --out over an existing spec (the .bak-N copy runs before the
    wrapped open), and apply --from <file> / apply --from <dir> when the
    destination's parent cannot be created, including the tree branch's per-entry
    writes and backups (#994).