From a874db8a4f372e13f568a3124a5641e8e1f07cb2 Mon Sep 17 00:00:00 2001 From: Birfy <15378933+Birfy@users.noreply.github.com> Date: Fri, 20 Nov 2026 15:49:49 +0800 Subject: [PATCH] release: v0.11.0 --- .claude-plugin/marketplace.json | 2 +- CHANGELOG.md | 609 +++++++++++++++++- README.md | 2 +- agentdescent/__init__.py | 2 +- changelog.d/1005-efficient-artifact-search.md | 7 - changelog.d/1008-dest-link.md | 6 - changelog.d/1011-evoskill-learned-line.md | 5 - changelog.d/1015-skill-bank-store-leaf.md | 8 - changelog.d/1017-demo-dest-link.md | 7 - changelog.d/1018-algorithm-spec-allowlist.md | 7 - changelog.d/1029-surface-out-link.md | 7 - changelog.d/1030-stores-symlink-leaf.md | 7 - changelog.d/1032-surface-error-envelope.md | 5 - changelog.d/1039-gepa-selection-warning.md | 5 - changelog.d/1041-ports-leaf-guard.md | 9 - changelog.d/1043-capability-trace-leaf.md | 11 - changelog.d/1045-ports-nondir-leaf.md | 6 - .../1046-capability-writer-durability.md | 7 - changelog.d/1048-plan-probe.md | 8 - changelog.d/1053-governed-integrity.md | 6 - changelog.d/1063-span-boundary-redaction.md | 7 - changelog.d/1064-experience-store-leaf.md | 5 - changelog.d/1070-secret-key-components.md | 9 - changelog.d/1082-cli-write-dir-enforcement.md | 8 - changelog.d/1082-host-env-scrub.md | 13 - changelog.d/1083-otel-wire-path.md | 14 - changelog.d/1085-cost-card-fallback.md | 8 - changelog.d/1089-noise-refusal-cause.md | 6 - changelog.d/1091-negative-gain-note.md | 4 - changelog.d/1092-resume-round-numbering.md | 7 - changelog.d/1093-probe-failed-next.md | 5 - changelog.d/1094-mcp-plan-probe.md | 9 - changelog.d/1095-credential-word-suffix.md | 9 - changelog.d/1096-handoff-failed-child.md | 11 - changelog.d/1098-env-scrub-exception.md | 6 - changelog.d/1100-manifest-secret-keys.md | 10 - changelog.d/1101-handoff-run-epoch.md | 7 - changelog.d/1106-algorithms-run-hint.md | 7 - .../1107-demo-tool-description-evoskill.md | 5 - .../1108-algorithm-spec-default-out.md | 7 - changelog.d/1109-store-quote.md | 6 - changelog.d/1112-lab-grouping.md | 6 - .../1113-plan-headroom-artifact-shape.md | 10 - changelog.d/1118-apply-created-target.md | 10 - changelog.d/1119-wrapped-trace-digest.md | 4 - .../1120-publisher-fast-forward-docs.md | 4 - changelog.d/1122-algorithm-spec-target-dir.md | 8 - changelog.d/1124-text-apply-missing-parent.md | 5 - changelog.d/1130-star-history.md | 3 - changelog.d/1131-docs-first-run-nav.md | 4 - changelog.d/1132-cost-page-pilot-shipped.md | 4 - changelog.d/1137-pilot-multi-seed-artifact.md | 4 - changelog.d/1138-artifact-snapshot.md | 6 - changelog.d/1141-pilot-next-artifact.md | 6 - .../1145-demo-algorithm-next-evolve.md | 6 - changelog.d/1146-first-run-transcript.md | 5 - changelog.d/1149-snapshot-apply-text.md | 2 - changelog.d/1150-snapshot-artifact-id.md | 8 - changelog.d/1154-snapshot-eval-spec-json.md | 3 - .../1155-baseline-sample-failed-envelope.md | 7 - ...1157-baseline-sample-quote-actual-tasks.md | 5 - changelog.d/1161-claim-expiry-refs.md | 5 - changelog.d/1164-snapshot-eval-skill-name.md | 6 - changelog.d/1166-run-call-cap.md | 11 - changelog.d/1171-lab-hides-internal-run.md | 4 - changelog.d/1174-help-epilog.md | 7 - changelog.d/912-mcp-http-bearer-token.md | 6 - changelog.d/920-embedding-facade.md | 8 - changelog.d/923-histogram-method.md | 6 - changelog.d/923-otlp-exporter.md | 12 - changelog.d/924-artifact-store-protocol.md | 4 - .../924-evaluation-provider-protocol.md | 6 - changelog.d/928-superseding-report.md | 7 - changelog.d/937-pilot-core.md | 10 - changelog.d/938-falsification-acceptance.md | 16 - .../940-verifier-admission-universality.md | 7 - changelog.d/940-verifier-admission.md | 11 - changelog.d/944-noise-calibration.md | 5 - changelog.d/945-cost-per-accepted-diff.md | 13 - changelog.d/946-failure-stratified-sampler.md | 7 - changelog.d/947-memory-spans.md | 7 - changelog.d/949-harness-integrity.md | 8 - changelog.d/951-evidence-card-1042.md | 6 - changelog.d/951-evidence-card-check.md | 5 - changelog.d/951-evidence-cards-headline.md | 6 - changelog.d/951-evidence-cards.md | 11 - changelog.d/962-dsh-super-agent.md | 6 - changelog.d/963-nested-run-refusal.md | 7 - changelog.d/964-long-run-handoff.md | 6 - docs/assets/experience.js | 2 +- docs/docs.md | 2 +- docs/index.md | 6 +- docs/install.md | 4 +- docs/manifest.md | 4 +- docs/plugin-quickstart.md | 4 +- docs/plugins.md | 2 +- docs/start.md | 2 +- .../claude-code/.claude-plugin/plugin.json | 2 +- integrations/dsh-agentdescent/package.json | 2 +- mkdocs.yml | 2 +- 100 files changed, 627 insertions(+), 617 deletions(-) delete mode 100644 changelog.d/1005-efficient-artifact-search.md delete mode 100644 changelog.d/1008-dest-link.md delete mode 100644 changelog.d/1011-evoskill-learned-line.md delete mode 100644 changelog.d/1015-skill-bank-store-leaf.md delete mode 100644 changelog.d/1017-demo-dest-link.md delete mode 100644 changelog.d/1018-algorithm-spec-allowlist.md delete mode 100644 changelog.d/1029-surface-out-link.md delete mode 100644 changelog.d/1030-stores-symlink-leaf.md delete mode 100644 changelog.d/1032-surface-error-envelope.md delete mode 100644 changelog.d/1039-gepa-selection-warning.md delete mode 100644 changelog.d/1041-ports-leaf-guard.md delete mode 100644 changelog.d/1043-capability-trace-leaf.md delete mode 100644 changelog.d/1045-ports-nondir-leaf.md delete mode 100644 changelog.d/1046-capability-writer-durability.md delete mode 100644 changelog.d/1048-plan-probe.md delete mode 100644 changelog.d/1053-governed-integrity.md delete mode 100644 changelog.d/1063-span-boundary-redaction.md delete mode 100644 changelog.d/1064-experience-store-leaf.md delete mode 100644 changelog.d/1070-secret-key-components.md delete mode 100644 changelog.d/1082-cli-write-dir-enforcement.md delete mode 100644 changelog.d/1082-host-env-scrub.md delete mode 100644 changelog.d/1083-otel-wire-path.md delete mode 100644 changelog.d/1085-cost-card-fallback.md delete mode 100644 changelog.d/1089-noise-refusal-cause.md delete mode 100644 changelog.d/1091-negative-gain-note.md delete mode 100644 changelog.d/1092-resume-round-numbering.md delete mode 100644 changelog.d/1093-probe-failed-next.md delete mode 100644 changelog.d/1094-mcp-plan-probe.md delete mode 100644 changelog.d/1095-credential-word-suffix.md delete mode 100644 changelog.d/1096-handoff-failed-child.md delete mode 100644 changelog.d/1098-env-scrub-exception.md delete mode 100644 changelog.d/1100-manifest-secret-keys.md delete mode 100644 changelog.d/1101-handoff-run-epoch.md delete mode 100644 changelog.d/1106-algorithms-run-hint.md delete mode 100644 changelog.d/1107-demo-tool-description-evoskill.md delete mode 100644 changelog.d/1108-algorithm-spec-default-out.md delete mode 100644 changelog.d/1109-store-quote.md delete mode 100644 changelog.d/1112-lab-grouping.md delete mode 100644 changelog.d/1113-plan-headroom-artifact-shape.md delete mode 100644 changelog.d/1118-apply-created-target.md delete mode 100644 changelog.d/1119-wrapped-trace-digest.md delete mode 100644 changelog.d/1120-publisher-fast-forward-docs.md delete mode 100644 changelog.d/1122-algorithm-spec-target-dir.md delete mode 100644 changelog.d/1124-text-apply-missing-parent.md delete mode 100644 changelog.d/1130-star-history.md delete mode 100644 changelog.d/1131-docs-first-run-nav.md delete mode 100644 changelog.d/1132-cost-page-pilot-shipped.md delete mode 100644 changelog.d/1137-pilot-multi-seed-artifact.md delete mode 100644 changelog.d/1138-artifact-snapshot.md delete mode 100644 changelog.d/1141-pilot-next-artifact.md delete mode 100644 changelog.d/1145-demo-algorithm-next-evolve.md delete mode 100644 changelog.d/1146-first-run-transcript.md delete mode 100644 changelog.d/1149-snapshot-apply-text.md delete mode 100644 changelog.d/1150-snapshot-artifact-id.md delete mode 100644 changelog.d/1154-snapshot-eval-spec-json.md delete mode 100644 changelog.d/1155-baseline-sample-failed-envelope.md delete mode 100644 changelog.d/1157-baseline-sample-quote-actual-tasks.md delete mode 100644 changelog.d/1161-claim-expiry-refs.md delete mode 100644 changelog.d/1164-snapshot-eval-skill-name.md delete mode 100644 changelog.d/1166-run-call-cap.md delete mode 100644 changelog.d/1171-lab-hides-internal-run.md delete mode 100644 changelog.d/1174-help-epilog.md delete mode 100644 changelog.d/912-mcp-http-bearer-token.md delete mode 100644 changelog.d/920-embedding-facade.md delete mode 100644 changelog.d/923-histogram-method.md delete mode 100644 changelog.d/923-otlp-exporter.md delete mode 100644 changelog.d/924-artifact-store-protocol.md delete mode 100644 changelog.d/924-evaluation-provider-protocol.md delete mode 100644 changelog.d/928-superseding-report.md delete mode 100644 changelog.d/937-pilot-core.md delete mode 100644 changelog.d/938-falsification-acceptance.md delete mode 100644 changelog.d/940-verifier-admission-universality.md delete mode 100644 changelog.d/940-verifier-admission.md delete mode 100644 changelog.d/944-noise-calibration.md delete mode 100644 changelog.d/945-cost-per-accepted-diff.md delete mode 100644 changelog.d/946-failure-stratified-sampler.md delete mode 100644 changelog.d/947-memory-spans.md delete mode 100644 changelog.d/949-harness-integrity.md delete mode 100644 changelog.d/951-evidence-card-1042.md delete mode 100644 changelog.d/951-evidence-card-check.md delete mode 100644 changelog.d/951-evidence-cards-headline.md delete mode 100644 changelog.d/951-evidence-cards.md delete mode 100644 changelog.d/962-dsh-super-agent.md delete mode 100644 changelog.d/963-nested-run-refusal.md delete mode 100644 changelog.d/964-long-run-handoff.md diff --git a/.claude-plugin/marketplace.json b/.claude-plugin/marketplace.json index 879605b2..39b5215d 100644 --- a/.claude-plugin/marketplace.json +++ b/.claude-plugin/marketplace.json @@ -11,7 +11,7 @@ "name": "agentdescent", "source": "./integrations/claude-code", "description": "Evolve skills, agents, prompts, code and plugins against examples with a parallel, merge-based optimiser.", - "version": "0.10.0" + "version": "0.11.0" } ] } diff --git a/CHANGELOG.md b/CHANGELOG.md index 01c92633..dc2771c7 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -6,6 +6,612 @@ All notable changes to AgentDescent are documented here. The format follows ## [Unreleased] +## [0.11.0] — 2026-11-20 + +### Added + +- Add an opt-in `ArtifactSearch` policy bundle that routes proposal generation + across configured models, screens duplicate candidates before held-out + evaluation, and feeds attributable measured gains back to the router. + - Preserve optional proposal identities through synchronous and asynchronous + evolution while keeping existing string proposers compatible. + +- `agentdescent plan --probe` and `agentdescent algorithms run … --probe` send one + minimal model call per model-backed agent and reflector before a run exists, so + an expired subscription, an exhausted quota, a wrong model name or a dead + endpoint is refused there with the same `SubscriptionInactive` / + `QuotaExhausted` / `ProviderConfigError` answer `evolve` would have given — and + without a `failed round 0/1` entry in the run store. Off by default: plain + `plan` still sends nothing (#1048). + +- An `otel-wire` CI leg installs the `[otel]` extra and exports one span and one + metric to a closed port on `127.0.0.1`, asserting a refused connection is logged + and contained rather than raised out of `export_run`. A test pins that + `export_span` copies `ExportedSpan.attributes` verbatim — it filters and invents + nothing — so the no-content guarantee is `export_run`'s and the caller's for + spans built by hand (#1083). + + The leg also keeps the fail-fast row real where the extra is installed: the SDK + import is halted through the interpreter's own `sys.modules` marker instead of a + probe-and-skip, so construction still has to name the install line, and the + loopback endpoint rows assert the accepted endpoint on a stubbed seam rather than + building a live provider. `docs/metrics.md` now says credentials go in + `OTEL_EXPORTER_OTLP_HEADERS`, never in the URL (#1083). + +- MCP `plan` takes `probe: bool = False` and, when true, runs the same + preflight as `plan --probe`: one minimal call per model-backed agent, with the + rows under `probe` in the result. A refusal makes the plan `ok: false` and + points `next` at the repair action instead of `start`, so a host that follows + `next` no longer discovers an expired subscription only after the run exists. + The default is unchanged: plain `plan` still sends nothing and returns its old + key set. The shared host skill teaches the probe step, so every installed + host now asks for it on the first use of a model (#1094). + +- `plan` can now catch two wasteful runs before they are paid for. The opt-in + `--baseline-sample N` scores the seed artifact on the first N tasks (one agent + call per task, quoted before the first call) and a seed at or above + `--saturation` (default `0.90`) prints `no headroom` and withholds the paid + run's quote unless `--force` is passed; `--json` carries the result as + `baseline` and sets `ready: false`. A plain `plan` still sends nothing. + `plan` and the pilot report also say when the artifact holds one key (a + `SingleSlot` prompt, a one-file tree): worker proposals contradict on that key, + so extra workers buy best-of-N selection, not merging (#1113). + +- A Star History chart at the end of the README, so visitors can see the + repository's GitHub star growth over time (#1130). + +- Portable artifact snapshots capture a complete FileTree candidate + (editable files plus frozen context) and the ledger's existing + id/parent/root lineage, and can be re-scored on another compatible + dataset without further search (`agentdescent snapshot + export|show|eval|apply`, #1138). + +- A run-level call cap that holds before the call (`--max-calls`, `--max-usd`, + also `max_calls` / `max_usd` in a spec's `evolve` block) (#1166). Unlike + `max_calls` on `evolve()`, which is a round barrier and lets a four-worker round + overshoot, the guarded cap is charged before each model call: the call that + would cross the limit is not made, the round closes without a partial merge or a + traceback, and the run ends `done` with `stop_reason=budget` and a + `stop_detail` naming the limit and the observed value, in `status.json` and in + `agentdescent status`. `0` is a real limit (no calls); a negative or non-numeric + value is refused at validation with the field and the value named; with neither + flag set the run and `status.json` are unchanged. Offline, no new dependency. + +- `agentdescent.facade` is the versioned import surface for embedding the engine: + a curated `__all__` covering Path A and the durable `LoopKernel` path, with + `API_VERSION = "1"` and a promise that changes are additive within a version and + removals need a bump. `tests/test_facade.py` pins the list, so widening or + narrowing it is a deliberate edit rather than an accidental import; nothing + outside the list — including the package root's research exports — is covered + (#920). + +- An opt-in OpenTelemetry bridge (`agentdescent.otel_export`): `export_run()` writes + one span per round and one per decision, and re-publishes the `rsi_metrics` + families plus a round-reward histogram. It is off unless + `OTEL_EXPORTER_OTLP_ENDPOINT` is set, a host can inject any object with + `export_span`/`export_metric` (so tests need no collector), and the new optional + `agentdescent[otel]` extra is imported only when an `OtlpExporter` is + constructed — importing the package and running normally needs nothing new. No + prompts or responses are exported by default. An endpoint carrying + credentials is refused without the value being echoed, and plain + `http://` is accepted only for a literal loopback collector; a remote + collector needs TLS terminated in front of it (#923). + +- `agentdescent.ports` adds the `ArtifactStore` protocol and the + `FileArtifactStore` reference backend over a content-addressed directory tree; + an id, revision, parent or path that could leave the root is refused (#1026). + +- `agentdescent.ports` adds the `EvaluationProvider` protocol and the + `CsvEvaluationProvider` reference backend, so a split and the scorer a run + optimises can come from a database or an internal judge service instead of a + file; `oracle` and `judge` report `None` rather than a score they cannot measure + (#924). + +- `agentdescent.pilot` turns one evolution run into one report a user can defend: + the seed and the result are scored on the same disjoint test tasks, the paired + scores become `TrustEvidence`, and `promote()` returns `promote` / `defer` / + `reject` with the reasons, cost, manifest and next action. A test task that + raised is run-health evidence rather than a zero, an unchanged artifact reports + `noise_only`, a missing contamination check defers, and a contaminated dataset + rejects before any model call. `pilot_scores()` exposes the same decision seam + for a run that already happened elsewhere; the CLI and `apply` are separate + follow-ups (#937). + +- Falsification-gated lesson admission: `FalsificationAcceptance` fills the existing + `Policies(acceptance=...)` slot and decides a commit on per-task measurements + instead of held-out counts alone. A lesson commits only with a measured gain on the + target-transfer partition, no per-task regression on the archived partition, and no + verdict change on the falsification partition; each reject reason + (`no-target-transfer`, `archived-regression`, `falsification-moved`) names the + failing task ids. Probes are caller-supplied (`MergeContext.probes`), and when they + are unavailable the policy defers (`AcceptDecision(defer=True)`) rather than falling + back to the count-only gate; all three partitions are required, and an empty one is + refused at construction instead of deferring on every candidate. A deferral under a + category this build does not know reports `probes-unavailable` rather than raising + inside the merge thread. A committed lesson hands its review evidence + (`rule_id -> {precondition, behavior_change, expected_effect, verifier, trace_ids, + evidence_card_hash}`) to the shared provenance writer as + `AcceptDecision.provenance` (#938). + +- `VerifierAgreement.admission_report()` now measures the cross-prompt error + overlap of the accepted stratum: the share of audit prompts on which a commit + the verifier accepted is oracle-denied, reported as `repeat_overlap`, with + `high_universality` at half the prompts over three audited accepts. A high + overlap blocks the admit the mean gains would allow; an unmeasurable one (a + single prompt, or a thin sample) stays `None` and never reads as zero (#940). + +- Verifier admission on both strata: `VerifierAgreement(rejected_sample_rate=...)` + audits a sample of the commits it rejected, and `admission_report()` reports + precision on the accepted stratum with its bootstrap interval and + `rejection_accuracy_{lower,upper}`, the interval on the share of rejections that + were correct, so a verifier that rejects good work no longer scores 1.0 on + precision alone. A demote needs a bound already under 0.5 on a stratum with at + least three audited decisions, an admit needs both pessimistic bounds at 0.8, and + a thin sample defers. `require_verifier_admission=True` carries it on + `EnsembleReport.admission`: missing, incomplete or insufficient defers, `reject` + fails closed (#940). + +- Baseline noise calibration repeats the baseline and reports a `noise_floor`, + task/model/run variance decomposition, `gain_over_noise` and `distinguishable`; + a gain inside the floor is labelled "not distinguishable from noise" and is + never promotion evidence (#944). + +- A per-diff `Cost` block (calls, tokens, wall seconds, optional USD), the + `cost_efficiency` / `cost_status` reads, a `CostPolicy` acceptance wrapper that + refuses a quality gain whose cost efficiency regressed (and defers when cost is + required but unmeasured), a `cost_report` with cost per accepted diff and per + point of gain, and `PromotionRules(require_cost=True)` (#945). A report average + is `None` unless every decision measured the unit being divided -- a row priced + in tokens only cannot make the dollar figure look cheaper than the evidence. + This also fixes the API reference for the name it exports: `docs/api.md` was + generated from a flat short-name map, so `agentdescent.Cost` was shadowed by + `agentdescent.audit.scorecard.Cost` and the page documented a signature the + top-level import does not return. The generator now keys by `(module, name)` and + `agentdescent.cost` has its own section, so both types are documented. + +- `FailureStratifiedSampler` stratifies tasks by base failure signature, keeps + search/select/final sets disjoint, and reports stratum coverage, overlap and + transfer status so one dominant failure mode cannot take over the search set + (#946). It is exported from the package root, the transfer comparison keeps a + float floor so equal means report `holds`, and the pick refusal names the + supported wiring (`tasks=sampler.split().search`). + +- Memory operation telemetry: `memory.read` / `memory.write` / `memory.update` / + `memory.forget` spans with lesson provenance (precondition, evidence hash, + source trace ids, verifier, expected effect), plus a memory audit block on the + telemetry and experience reports that counts each operation, flags lessons + without provenance, and lists forgotten or rolled-back lessons with a reason and + attributable regressions (#947). + +- `HarnessIntegrity` checks a plugin/harness before/after snapshot before + promotion: guardrail enforcement-point order (policy before tool mediation), + permission and allow-rule state, the action set unless the diff declares exactly + that change, and the decision-path record for every enforced relationship. + `govern_harness_change(..., integrity=...)` rejects a broken composition before + the proof, approval and promotion seams, and the report names each broken + relationship with the state it observed (#949). + +- `reports/evidence/1042-ports-leaf-write-through.json` publishes the + leaf-write-through fix as a v1 evidence card — `code` and `reliability` cite + the merged sha, the guarded lines and the regression test, and the other ten + fields say `not reviewed` — with a test that keeps every published card + schema-valid (#951). + +- `python -m tools.evidence_card --check` validates the 12-field evidence card + schema v1 over the named `*.json` cards, and refuses a missing, unknown or blank + field by name; `not reviewed` is the explicit way to say a field has no evidence + (#1025). + +- Two evidence cards for headline results from `reports/` -- the audit's constant + rubric tying the judge at 0.836, and the 177-pair verifier bias with its 0.2203 + resolution floor -- plus `reports/evidence/README.md`, which states that a new + claim needs a card, and tests pinning each card's number, its reproduction and + the rule itself (#951). + +- `docs/dsh-super-agent.md` and `scripts/verify-dsh-superagent.sh`: the zero-code + composition of AgentDescent with dsh-super-agent — the two Cordis rows in the + `superagent` profile, the skill-or-MCP fallback, the `plan → start → status → + show` worker loop with no auto-apply, write-dir and budget settings for a + worker, and a gated check that skips cleanly without a live host (#962). + +- A worker session (`AGENTDESCENT_NESTED=1`, set in every host worker the runner + spawns) can no longer start a nested run through the CLI: `agentdescent evolve` + refuses with exit 2 and `reason: nested-run` before it reads the spec or creates + a run, and `doctor` reports `nested_guard` instead of only a boolean. The MCP + `start` stub carries the same reason and sentence, so a host branches on one + contract. Offline, no new sandbox or auth layer (#963). + +- `scripts/dsh-superagent-run.sh` hands a long evolution run off a + dsh-superagent worker: it starts the run in a detached `tmux` session + (`setsid` fallback), writes `waiting.json` with the run id, round and run + directory, and returns at once, so a worker turn no longer holds its slot for + the length of the run (#964). + +### Changed + +- A product-surface refusal prints one `error: ` line — the same envelope + the `cannot write` writers use — instead of prefixing `SurfaceError:`, and the + `--json` `error` field drops the prefix too. Any other exception keeps its class + name (#1032). + +- `agentdescent governed-harness` accepts a caller-supplied `integrity` pair and + passes it to `govern_harness_change`, so a spec whose after-state reorders the + guardrail path, drops an allow rule, loses a decision record or changes the + action set without declaring it is rejected before proof, approval and + promotion instead of silently skipping the composition check (#1053). + +- `agentdescent apply` honours `AGENTDESCENT_SURFACE_WRITE_DIRS`: a destination, + a materialised tree or a `.bak-N` slot outside the allowlist is refused with + exit 2 before anything touches disk. Inside a host worker + (`AGENTDESCENT_NESTED=1`) the allowlist defaults to the run's cwd plus + `$AGENTDESCENT_HOME`; outside one it is only enforced when the variable is set, + so a terminal `apply --to ` behaves as before. `show --diff` still + renders and reports the refused target in `apply_plan` (#1082). + +- `docs/cost.md` and the `Cost.from_cards` / module docstrings now state what the + card fallback actually returns in-tree: no producer writes + `cost_tokens`/`cost_wallclock` onto an `EvidenceCard`, so `Cost.from_cards` + yields `Cost()` — unknown, not a free diff — and a caller that wants the gate to + price a diff passes `MergeContext.diff_cost = Cost.from_usage(usage)`. The page + also warns that `require_cost=True` defers every candidate on a run whose cards + carry no cost (#1085). + +- `agentdescent --help` now lists the first-run path (`demo`, `init`, `plan`, + `evolve`, `show`, `apply`, `doctor`, `status`) plus `mcp`, and every other verb + is grouped under a new `lab` subcommand (`agentdescent lab campaign …`); each + verb keeps its old top-level path, its arguments, its output and its exit code, + and the same declaration mounts both paths (#1112). + +- The README and documentation landing pages now lead with one first-run path, + and the docs navigation is grouped by getting started, hosts and plugins, and + algorithms, instead of one long mixed list (#1131). + +- The top-level `agentdescent --help` now ends with a short next-steps block + naming `agentdescent install `, `agentdescent algorithms list` and + `agentdescent lab --help`, so the two verbs the first-run docs use are + discoverable on the first screen without reading the ~40 verbs under `lab` + (#1174). The listing, the usage line and `DEFAULT_SURFACE` are unchanged from + #1112, and no verb, argument, output or exit code changed. + +- The results and audit pages now point at evidence cards for their headline + claims: one card per claim under `reports/evidence/`, each answering all twelve + fields and writing `not reviewed` where nobody measured. Eight new cards cover + the one-call path, bringing your own agent, the efficiency rows, equal-budget + HotpotQA, the merge cost lever, judge bias across five workloads, ground-truth + repair, and the rule search's residual against the sigma floor. The directory + indexes them in `reports/evidence/README.md`, the two bench cards are recomputed + in CI from the committed runs, and the rest say on the card that their number + came from a live run — so the gap between "we measured it" and "you can check + it" is visible where the claim is (#951). + +### Fixed + +- `init` and `algorithms run` refuse a symlink planted at the destination — the + `--out` leaf, the `.opencode-providers.json` file, and the starter + directories a skill library is seeded into — instead of writing through it, so + the spec can no longer land outside the tree and no `.bak-N` copy holds a file + from outside it (#1008). + +- `agentdescent demo --algorithm evoskill` now says what it learned in one + readable line: each induced `SKILL.md` is named with its rule bullets instead + of a whole escaped file, and a seeded `rules.md` the run left unchanged is no + longer listed as learned (#1011). + +- `agentdescent skills` now treats the store leaf as a regular file: a directory, + device or FIFO named by `--store` is refused with `cannot read skill bank …: not + a regular file` instead of reporting `skills: 0`, and every write refuses a + symlink store before it creates anything, so `--apply` can no longer replace the + link and leave the link and its target disagreeing. A link to a file still + reads. `audit --apply` now prints `applied: false` when no status changed (it + used to echo the flag), which is a visible change to that output line. + +- `agentdescent demo --dir` now refuses a symlink planted at a fixture leaf or at + `spec.json` instead of following it: the run exits `2` with + `cannot write …: '' is a symlink` before the first byte, so the demo's + bytes can no longer land outside `--dir` or overwrite the file the link points + at. The check is `filetree.dest_problem`, the same rule `init` (#1008) and the + tree writer (#977) use, and a link above `--dir` is still allowed (#1017). + +- The MCP `algorithm_spec` tool now refuses an `out` path outside the operator's + `AGENTDESCENT_SURFACE_WRITE_DIRS` allowlist instead of writing the spec there. + It was the only writing tool that skipped `_check_write`, so a host could name + any path on the server's filesystem; the default allowlist is unchanged (the + server's cwd and `$AGENTDESCENT_HOME`), and an allowed `out` plans as before + (#1018). + +- The product surfaces (`governed-harness`, `platform`, `capability`, + `meta-campaign`, `ops`, `tenant`, `traces`, `compile-skill`) and the MCP writers + now refuse a symlink planted at the output leaf — one `error: SurfaceError: + cannot write …: '' is a symlink` line, exit 2, without touching the link + or its target — instead of silently replacing the operator's link with a + regular file (#1029). + +- The workspace registry, the consolidation archive (`archive`, `mark_exam`) and + the run store's atomic JSON writer refuse a symlink planted at their own file + instead of silently replacing it with a regular file: `cannot write : + '' is a symlink`, raised before any temporary exists, with the link and + its target's bytes untouched. A symlinked directory above the file still + works (#1030). + +- `plan` no longer warns that `n_workers` is "best-of-N selection" for + `algorithm: gepa`: its Pareto pool admits every surviving proposal, so the + "collapse to one candidate / name a `reflect` model" advice did not apply on + the aggregator_factory path (#1039). + +- `FileArtifactStore` no longer writes through a link planted at a leaf it + writes: `promote` refuses `` before it reads or removes through it, and + `_commit` refuses `.json` before staging, both as `ValueError: cannot + write : '' is a symlink`. The sidecar is now staged and renamed + over that leaf, so a hardlink at `.json` is replaced instead of + truncating the inode it shares with bytes outside `root`. The link and those + bytes are untouched, and a symlinked directory above the store still works + (#1041). + +- `agentdescent capability init --demo` refuses a symlink planted at + `capability-traces.jsonl` or at `--out` before the first byte, instead of + following the link and destroying the file it points at; a link at `--out` no + longer leaves a half-written traces fixture behind (#1043). + + The traces fixture is now staged in a temporary file and renamed into place, so + a hardlink at that leaf keeps the bytes it shares with the file outside, a link + planted between the guard and the write is never dereferenced, and a directory + at the leaf is refused with `cannot write : not a file` instead of a raw + `IsADirectoryError`. + +- `FileArtifactStore` now refuses a regular file where a revision tree belongs and + a directory where its sidecar belongs with `ValueError` before anything is + written, instead of the raw `NotADirectoryError`/`IsADirectoryError` from the + rename; a refused sidecar no longer leaves the revision listed without one + (#1045). + +- `agentdescent capability init` writes the spec before the demo traces, so an + `--out` that is an existing directory fails at exit 2 without leaving a + `capability-traces.jsonl` in its parent. The staged traces are flushed and + `fsync`ed before the rename, and a `KeyboardInterrupt` while they are written + removes the temporary file instead of leaving a `.staging-*.jsonl` behind + (#1046). + +- `ExperienceStore.add` refuses a link planted at the `--store` leaf before the + first byte — a symlink, a hardlink and a dangling symlink each appended + wherever the link pointed — and refuses a leaf that is not a regular file, + with the sibling writers' `cannot write : ` shape (#1064). + +- Telemetry exports lose no more than credentials: a key is dropped when a whole + component of it names a credential, not when it merely contains the substring + `token` or `secret`. `tokens`, `max_tokens`, `input_tokens`, `output_tokens`, + `token_count`, `tokens_in`, `budget_tokens`, `secret_free` and + `secrets_scanned` are written again (the cost and budget counters among them), + while `api_key`, `apikey`, `access_key`, `client_secret`, `private_key`, + `password`, `authorization`, `credential` and the `*_token` spellings still + drop (#1070). + +- A detached run no longer inherits the worker's own session: `runstore.launch` + drops the host-session markers it lists, except the four host config roots it + keeps (`CLAUDE_CONFIG_DIR`, `CODEX_HOME`, `DSH_HOME`, `XDG_CONFIG_HOME`), before + it starts the grandchild, so the host worker's session identity and MCP server + configuration stay with the host while the credentials and `PATH` the run needs + still pass through (#1082). The marker list is `SESSION_MARKERS` (`DSH_*`, + `CLAUDE*`, `CODEX*`, `MCP_*`, the OpenCode config carriers and + `XDG_CONFIG_HOME`), the one the candidate worker already dropped + (`agentdescent.agents.worker_env`); it used to be a second, shorter copy, and + five of its eleven entries crossed the scrub. Those four roots name + configuration rather than session, and the run resolves a host's real home from + them (#1082). + +- Trust evidence now names the actual reason a gain is refused as noise: too few + baseline repeats to bound a floor, a gain inside the floor (printed as a band + around the baseline mean rather than an absolute interval against a delta), or + a gain below `min_effect`. `gain_over_noise` is `0.0` when the floor is + undefined instead of a ratio over an epsilon (#1089). + +- `trust_evidence` now explains a refused gain that is zero or negative as "the + candidate did not improve on the baseline" instead of blaming the noise floor, + and the note's docstring lists all four causes in code order (#1091). + +- A resumed run continues its ledger's round numbering instead of restarting at + `0`, and `status.best_reward` no longer rewinds below a reward the run's own + history still holds. `execute()` passes one past the last recorded round as the + new `round_base` and seeds the best from the existing rows; a checkpoint further + ahead than the ledger still wins, and an empty ledger still starts at `0` + (#1092). + +- A failed `plan --probe` or `algorithms run --probe` no longer ends with + `next: agentdescent evolve`. The last line (and the JSON `next`) is the + repair action, `--json` sets `ready: false` while leaving `ok` about the + plan itself, and a successful or omitted probe is unchanged (#1093). + +- `_is_secret_key` no longer lets a credential through when it arrives as one + word: `PGPASSWORD`, `s3secretkey` and `AWSSecretAccessKey` are matched by + whole-word suffix on a single-component key, and the adjacent-pair lookup + tolerates one plural `s`, so `client_secrets` and `api_keys` are dropped too. + The plural spellings (`passwords`, `pgpasswords`, `db_secrets`) match in the + last position or as a whole component. The `#1070` counters (`tokens`, + `max_tokens`, `secrets_scanned`) still survive, and `auth` stays a recorded + non-goal (#1095). + +- The handoff wrapper no longer reports a run that never started as a successful + handoff: when the wait window closes with the child gone and no new + `status.json` where the wrapper can see it, it writes `waiting.json` with + `status: failed` and `state: exited`, and exits nonzero with the child's own + reason left in `worker.log`. A child that is still running and has not written + its status yet is unchanged: `waiting`, `found: false`, exit 0. A store the + caller owns as the run directory is still reported from its own rewritten + `status.json`. The verdict reads the store again after the liveness probe, so a + status written while the probe ran is still this handoff's. The page says all + three (#1096). + +- The env-scrub note no longer says it drops "every" session marker and then keeps + some of them: `docs/dsh-super-agent.md` and the #1082 release note now name the + four host config roots (`CLAUDE_CONFIG_DIR`, `CODEX_HOME`, `DSH_HOME`, + `XDG_CONFIG_HOME`) as an exception to the marker list before that list appears, + so the first clause cannot be read as dropping them (#1098). + +- `manifest.redact` now matches every credential spelling `rsi_telemetry` does: + `PGPASSWORD`, `s3secretkey`, `client_secrets` and the other one-word or plural + forms are redacted before hashing and before a manifest or trajectory record is + written, instead of being stored and hashed in clear. The manifest predicate + keeps its own rules (`bearer`, and the `*_key` suffix that `sort_key` relies on) + and delegates the rest, so a key redacted before still is. Records whose + `model_config` or `extra` carried one of the newly matched spellings get a new + `model_hash` / `spec_hash`, so `verify_manifest` reports those inputs as changed + for them: that is the fix redacting what it used to leak, not a bug (#1100). + +- The `scripts/dsh-superagent-run.sh` handoff no longer reports an earlier run that + is still going. The launch epoch is taken with the `status.json` snapshot, and a + changed file is only reported when its `started` is numeric and no older than + that epoch, so a handoff into a reused store names its own run — or + `found: false` — instead of the earlier run, whose `run_id` the page's read-back + would then have polled (#1101). + +- `algorithms run` no longer prints the `--rounds 1 --workers 1` quick-trial hint + after a refused `--probe`: with the kind's default rounds and workers the hint + used to land below `next: renew the subscription …`, so the last line on screen + answered a question the refusal had already answered. The hint still prints + after a passing probe and when no probe runs, and every other line of the + refusal is unchanged (#1106). + +- The MCP `demo` tool description no longer says evoskill needs a real model: + it lists the same offline cards as `agentdescent demo --help` (including + evoskill, which defaults to the `skill_library` fixture) and accepts + `kind` `skill_library` (#1107). + +- `algorithm_spec` no longer defaults its starter spec to a path beside the + user's target, which the default write allowlist refused: an omitted `out` now + writes `agentdescent..spec.json` under `$AGENTDESCENT_HOME` (or the first + explicitly allowed root), so the shortest documented path plans on an untouched + install. The tool description and the shared `SKILL.md` say so, and a refused + explicit `out` names the allowed roots and an `out` that works (#1108). + +- `apply` into a tree target that no longer exists now says so: the plan carries + `"created": true` and a real run prints `created from run ; nothing + was restored and no backup was taken` once on stderr, instead of a + `written`/`backup: []` plan that read like an ordinary in-place update. An + in-place apply keeps `"created": false` and prints no line; so does a no-op one, + because an empty artifact into a missing target writes nothing and must not + claim a recreate. The dry run reports the flag without the line, since it + creates nothing. The flag is emitted for the tree path only: a text `--to` plan + is unchanged and carries no `created` key (#1118). + +- `traces digest --events FILE` now reads a wrapped `{"events": [...]}` object + (compact or pretty-printed) as the same digest as a JSON array, instead of + reporting a clean bill or exiting 2 (#1119). + +- The GitHub publisher page now documents fast-forward-only proposal branch + updates: a second publish of a `behind` or `diverged` branch is refused with + `refusing to rewrite it` instead of promising a force update (#1120). + +- `algorithm_spec` no longer writes its default starter spec inside a directory + target: when the only allowed write roots are the target itself or below it, + the call returns `ok: false` naming the target and + AGENTDESCENT_SURFACE_WRITE_DIRS instead of + dropping `agentdescent..spec.json` into the user's directory; a usable + root outside the target is still preferred, and the description no longer + claims `$AGENTDESCENT_HOME` when that home is the target (#1122). + +- A text-kind `apply --to /rule.md` now creates the missing parent + directory before writing, so the dry run it had already reported green is true: + the real run exits 0 and the file lands at that path. A genuinely unwritable + `--to` still exits 2 with the single `error: cannot write …` line (#1124). + +- The cost page no longer says `agentdescent/pilot.py` is absent: the module shipped + in #1121, its `PilotReport` carries the run's cost mapping, and `CostReport` + stays the cost-only surface that page describes (#1132). + +- `pilot` no longer reports the seed-0 artifact next to scores pooled + across every seed. A single artifact is reported only when every seed + produced it; otherwise the report summarizes per seed (#1137). + +- A promoting multi-seed pilot report no longer tells the reader to "review the + artifact diff" or "apply it" when the seeds produced different artifacts and the + report names none. The steps point at `run.per_seed` and ask for one seed's + artifact to be chosen, reviewed and applied as an explicit step; the single-seed + and agreeing-seed reports keep their current text (#1141). + +- `demo --algorithm NAME` no longer ends with `evolve agentdescent.NAME.spec.json`. + That file does not exist until `algorithms run` writes it, so following the + footer (or `--json` / MCP `next`) failed with exit 2. The footer still points + at `algorithms run NAME ...`, which plans and then prints `evolve` with the + absolute path it just wrote (#1145). + +- The first-screen demo transcripts (`README.md`, `docs/start.md`, + `docs/plugin-quickstart.md`, `docs/testing-guide.md`) now show the seed baseline + the CLI prints, `held-out reward: 0.000 -> 1.000`, instead of a bare `1.000`, and + a docs guard keeps those four blocks on the command's output (#1146). + +- `snapshot apply` on a text snapshot now writes the rendered prompt to the file named by `--to`, with the same backup, no-op, and parent-creation rules as `apply --from`. It previously materialised the strategy's internal rule store (hash-named files) into a directory, and an empty-rules snapshot reported a no-op instead of writing the seed prompt. A directory `--to` is refused. Tree-kind snapshots are unchanged (#1149). + +- A snapshot's `artifact_id` is validated as one path-safe component when the + snapshot is built, and both staging writes (`snapshot eval`, `snapshot apply`) + stage under a fixed name inside their temp directory instead of deriving a + path from the id: applying or scoring a snapshot JSON received from someone + else can no longer make the staging write land outside that temp directory. + A malformed id is reported as a clean CLI error (exit 2), not a traceback + (#1150). + +- The snapshot re-evaluation examples now use `dest.json`, which `load_spec` + can read, instead of `dest.yaml` (#1154). + +- `plan --baseline-sample` now keeps the plan envelope when seed scoring fails + on a provider, subscription, or endpoint error: exit 1 (not 2), `--json` + writes `ok: true`, `ready: false`, a `baseline` object with the failure, and + a `next` repair step (the same wording `plan --probe` already uses), and the + human form keeps the plan plus `next:` / `then:`. Input errors such as + `--baseline-sample 0` still exit 2. No run-store entry is written (#1155). + +- `plan --baseline-sample` now quotes the number of tasks the sample can + actually score, not the requested N: a 4-task spec asked for 50 is priced + at 4 calls, and that count matches `baseline.sample`. A request below the + task count is unchanged (#1157). + +- A claim no longer expires under a pull request that links its issue with + `Refs #N` or `Part of #N`: the controller now reads those references when it + decides whether the claim's TTL is still fed by a push, while only closing + keywords still make a PR an agent PR (#1161). + +- `snapshot eval` of a `skill_dir` snapshot now exposes the candidate under + the snapshot's `artifact_id` in the destination layout (for example + `.claude/skills/csv-total/`), so a demo export re-evaluated with `--spec` + scores the learned skill instead of 0 because the dest agent looked under + `.claude/skills/artifact/` (#1164). + +- `agentdescent lab --help` and the usage line on a `lab` argparse error no + longer list the internal `run` worker in the `{...}` metavar; the verb stays + reachable as `lab run` (#1171). + +- `export_metric()` sent histograms through the counter's `add`, so the first real + export of an `agentdescent_rsi_round_reward` raised `AttributeError` out of + `export_run`. Histograms now call the SDK's `Histogram.record()`, and the fake + SDK gives each instrument kind only its own method, so the same mistake fails in + the offline suite instead of on a host's first run (#923). + +- The review gate now orders comments and reviews together by time and lets the + newest report at a tip decide, so a same-tip follow-up with no `Verdict:` line + (a withdrawal, an added requirement, an open question) voids an earlier PASS + there instead of leaving it counting. A report that names another tip is still + judged against that tip, a newer PASS still counts, and the `Ran:` / `Not run:` + handling for the winning PASS is unchanged (#928). + +### Security + +- Span values are redacted at the boundary that carries them out of the process: + `set_attribute`, `add_event`, a span's initial `attributes`, a directly + constructed `TelemetrySpan`, a `MemoryOperation.to_span_attributes()` and the + `exception` event's message now pass through the same key/value filter, so a + credential in a precondition, verifier, event detail or raised error is written + as `[redacted]` (#1063). + +- The handoff wrapper no longer builds its launch command as a shell string: the + log path and the run's command reach `tmux`/`setsid` as arguments, so a `--store` + containing `'`, `;`, `$` or a space is used literally instead of closing the + quote and running the rest of the path as shell words outside the worker's + process group (#1109). + +- `agentdescent mcp --transport streamable-http` can now require an + `Authorization: Bearer` header (`--http-token-env NAME`, with the token read + from the environment and never from the command line). A request without it is + answered 401 before an MCP session is created, so a leaked tunnel URL no longer + reaches `apply` or `start`; `--allow-remote` is refused without a token (#912). + ## [0.10.0] — 2026-11-13 ### Added @@ -7639,7 +8245,8 @@ First public release on PyPI as **`agentdescent`**. discrete-space `Aggregator`, staleness policies, DP/TP/PP parallelism, layered governance, and the provider-agnostic `agentdescent.agents` completion layer. -[Unreleased]: https://github.com/ThinkFlowLab/agentdescent/compare/v0.10.0...HEAD +[Unreleased]: https://github.com/ThinkFlowLab/agentdescent/compare/v0.11.0...HEAD +[0.11.0]: https://github.com/ThinkFlowLab/agentdescent/compare/v0.10.0...v0.11.0 [0.10.0]: https://github.com/ThinkFlowLab/agentdescent/compare/v0.9.0...v0.10.0 [0.9.0]: https://github.com/ThinkFlowLab/agentdescent/compare/v0.8.0...v0.9.0 [0.8.0]: https://github.com/ThinkFlowLab/agentdescent/compare/v0.7.0...v0.8.0 diff --git a/README.md b/README.md index 13038eff..41a6464b 100644 --- a/README.md +++ b/README.md @@ -20,7 +20,7 @@ ## Install ```bash -pip install "agentdescent[mcp]>=0.10.0" +pip install "agentdescent[mcp]>=0.11.0" agentdescent demo # a complete evolution, offline, no API key, ~10 s agentdescent doctor # which agent CLIs and provider keys this machine has ``` diff --git a/agentdescent/__init__.py b/agentdescent/__init__.py index 8380feba..6478dbeb 100644 --- a/agentdescent/__init__.py +++ b/agentdescent/__init__.py @@ -553,7 +553,7 @@ shard_round_robin, ) -__version__ = "0.10.0" +__version__ = "0.11.0" __all__ = [ "Contract", diff --git a/changelog.d/1005-efficient-artifact-search.md b/changelog.d/1005-efficient-artifact-search.md deleted file mode 100644 index 55781c01..00000000 --- a/changelog.d/1005-efficient-artifact-search.md +++ /dev/null @@ -1,7 +0,0 @@ -Added - -- Add an opt-in `ArtifactSearch` policy bundle that routes proposal generation - across configured models, screens duplicate candidates before held-out - evaluation, and feeds attributable measured gains back to the router. -- Preserve optional proposal identities through synchronous and asynchronous - evolution while keeping existing string proposers compatible. diff --git a/changelog.d/1008-dest-link.md b/changelog.d/1008-dest-link.md deleted file mode 100644 index 477b6932..00000000 --- a/changelog.d/1008-dest-link.md +++ /dev/null @@ -1,6 +0,0 @@ -Fixed -`init` and `algorithms run` refuse a symlink planted at the destination — the -`--out` leaf, the `.opencode-providers.json` file, and the starter -directories a skill library is seeded into — instead of writing through it, so -the spec can no longer land outside the tree and no `.bak-N` copy holds a file -from outside it (#1008). diff --git a/changelog.d/1011-evoskill-learned-line.md b/changelog.d/1011-evoskill-learned-line.md deleted file mode 100644 index 9b8dad45..00000000 --- a/changelog.d/1011-evoskill-learned-line.md +++ /dev/null @@ -1,5 +0,0 @@ -Fixed -`agentdescent demo --algorithm evoskill` now says what it learned in one -readable line: each induced `SKILL.md` is named with its rule bullets instead -of a whole escaped file, and a seeded `rules.md` the run left unchanged is no -longer listed as learned (#1011). diff --git a/changelog.d/1015-skill-bank-store-leaf.md b/changelog.d/1015-skill-bank-store-leaf.md deleted file mode 100644 index 1159e304..00000000 --- a/changelog.d/1015-skill-bank-store-leaf.md +++ /dev/null @@ -1,8 +0,0 @@ -Fixed -`agentdescent skills` now treats the store leaf as a regular file: a directory, -device or FIFO named by `--store` is refused with `cannot read skill bank …: not -a regular file` instead of reporting `skills: 0`, and every write refuses a -symlink store before it creates anything, so `--apply` can no longer replace the -link and leave the link and its target disagreeing. A link to a file still -reads. `audit --apply` now prints `applied: false` when no status changed (it -used to echo the flag), which is a visible change to that output line. diff --git a/changelog.d/1017-demo-dest-link.md b/changelog.d/1017-demo-dest-link.md deleted file mode 100644 index 78cc94e2..00000000 --- a/changelog.d/1017-demo-dest-link.md +++ /dev/null @@ -1,7 +0,0 @@ -Fixed -`agentdescent demo --dir` now refuses a symlink planted at a fixture leaf or at -`spec.json` instead of following it: the run exits `2` with -`cannot write …: '' is a symlink` before the first byte, so the demo's -bytes can no longer land outside `--dir` or overwrite the file the link points -at. The check is `filetree.dest_problem`, the same rule `init` (#1008) and the -tree writer (#977) use, and a link above `--dir` is still allowed (#1017). diff --git a/changelog.d/1018-algorithm-spec-allowlist.md b/changelog.d/1018-algorithm-spec-allowlist.md deleted file mode 100644 index 404a81bb..00000000 --- a/changelog.d/1018-algorithm-spec-allowlist.md +++ /dev/null @@ -1,7 +0,0 @@ -Fixed -The MCP `algorithm_spec` tool now refuses an `out` path outside the operator's -`AGENTDESCENT_SURFACE_WRITE_DIRS` allowlist instead of writing the spec there. -It was the only writing tool that skipped `_check_write`, so a host could name -any path on the server's filesystem; the default allowlist is unchanged (the -server's cwd and `$AGENTDESCENT_HOME`), and an allowed `out` plans as before -(#1018). diff --git a/changelog.d/1029-surface-out-link.md b/changelog.d/1029-surface-out-link.md deleted file mode 100644 index 2875cbed..00000000 --- a/changelog.d/1029-surface-out-link.md +++ /dev/null @@ -1,7 +0,0 @@ -Fixed -The product surfaces (`governed-harness`, `platform`, `capability`, -`meta-campaign`, `ops`, `tenant`, `traces`, `compile-skill`) and the MCP writers -now refuse a symlink planted at the output leaf — one `error: SurfaceError: -cannot write …: '' is a symlink` line, exit 2, without touching the link -or its target — instead of silently replacing the operator's link with a -regular file (#1029). diff --git a/changelog.d/1030-stores-symlink-leaf.md b/changelog.d/1030-stores-symlink-leaf.md deleted file mode 100644 index e71d61a1..00000000 --- a/changelog.d/1030-stores-symlink-leaf.md +++ /dev/null @@ -1,7 +0,0 @@ -Fixed -The workspace registry, the consolidation archive (`archive`, `mark_exam`) and -the run store's atomic JSON writer refuse a symlink planted at their own file -instead of silently replacing it with a regular file: `cannot write : -'' is a symlink`, raised before any temporary exists, with the link and -its target's bytes untouched. A symlinked directory above the file still -works (#1030). diff --git a/changelog.d/1032-surface-error-envelope.md b/changelog.d/1032-surface-error-envelope.md deleted file mode 100644 index cf148aba..00000000 --- a/changelog.d/1032-surface-error-envelope.md +++ /dev/null @@ -1,5 +0,0 @@ -Changed -A product-surface refusal prints one `error: ` line — the same envelope -the `cannot write` writers use — instead of prefixing `SurfaceError:`, and the -`--json` `error` field drops the prefix too. Any other exception keeps its class -name (#1032). diff --git a/changelog.d/1039-gepa-selection-warning.md b/changelog.d/1039-gepa-selection-warning.md deleted file mode 100644 index a922a01c..00000000 --- a/changelog.d/1039-gepa-selection-warning.md +++ /dev/null @@ -1,5 +0,0 @@ -Fixed -`plan` no longer warns that `n_workers` is "best-of-N selection" for -`algorithm: gepa`: its Pareto pool admits every surviving proposal, so the -"collapse to one candidate / name a `reflect` model" advice did not apply on -the aggregator_factory path (#1039). diff --git a/changelog.d/1041-ports-leaf-guard.md b/changelog.d/1041-ports-leaf-guard.md deleted file mode 100644 index 427d99e4..00000000 --- a/changelog.d/1041-ports-leaf-guard.md +++ /dev/null @@ -1,9 +0,0 @@ -Fixed -`FileArtifactStore` no longer writes through a link planted at a leaf it -writes: `promote` refuses `` before it reads or removes through it, and -`_commit` refuses `.json` before staging, both as `ValueError: cannot -write : '' is a symlink`. The sidecar is now staged and renamed -over that leaf, so a hardlink at `.json` is replaced instead of -truncating the inode it shares with bytes outside `root`. The link and those -bytes are untouched, and a symlinked directory above the store still works -(#1041). diff --git a/changelog.d/1043-capability-trace-leaf.md b/changelog.d/1043-capability-trace-leaf.md deleted file mode 100644 index 21343b44..00000000 --- a/changelog.d/1043-capability-trace-leaf.md +++ /dev/null @@ -1,11 +0,0 @@ -Fixed -`agentdescent capability init --demo` refuses a symlink planted at -`capability-traces.jsonl` or at `--out` before the first byte, instead of -following the link and destroying the file it points at; a link at `--out` no -longer leaves a half-written traces fixture behind (#1043). - -The traces fixture is now staged in a temporary file and renamed into place, so -a hardlink at that leaf keeps the bytes it shares with the file outside, a link -planted between the guard and the write is never dereferenced, and a directory -at the leaf is refused with `cannot write : not a file` instead of a raw -`IsADirectoryError`. diff --git a/changelog.d/1045-ports-nondir-leaf.md b/changelog.d/1045-ports-nondir-leaf.md deleted file mode 100644 index fdc3d1c4..00000000 --- a/changelog.d/1045-ports-nondir-leaf.md +++ /dev/null @@ -1,6 +0,0 @@ -Fixed -`FileArtifactStore` now refuses a regular file where a revision tree belongs and -a directory where its sidecar belongs with `ValueError` before anything is -written, instead of the raw `NotADirectoryError`/`IsADirectoryError` from the -rename; a refused sidecar no longer leaves the revision listed without one -(#1045). diff --git a/changelog.d/1046-capability-writer-durability.md b/changelog.d/1046-capability-writer-durability.md deleted file mode 100644 index a11dd1fa..00000000 --- a/changelog.d/1046-capability-writer-durability.md +++ /dev/null @@ -1,7 +0,0 @@ -Fixed -`agentdescent capability init` writes the spec before the demo traces, so an -`--out` that is an existing directory fails at exit 2 without leaving a -`capability-traces.jsonl` in its parent. The staged traces are flushed and -`fsync`ed before the rename, and a `KeyboardInterrupt` while they are written -removes the temporary file instead of leaving a `.staging-*.jsonl` behind -(#1046). diff --git a/changelog.d/1048-plan-probe.md b/changelog.d/1048-plan-probe.md deleted file mode 100644 index b6a0c713..00000000 --- a/changelog.d/1048-plan-probe.md +++ /dev/null @@ -1,8 +0,0 @@ -Added -`agentdescent plan --probe` and `agentdescent algorithms run … --probe` send one -minimal model call per model-backed agent and reflector before a run exists, so -an expired subscription, an exhausted quota, a wrong model name or a dead -endpoint is refused there with the same `SubscriptionInactive` / -`QuotaExhausted` / `ProviderConfigError` answer `evolve` would have given — and -without a `failed round 0/1` entry in the run store. Off by default: plain -`plan` still sends nothing (#1048). diff --git a/changelog.d/1053-governed-integrity.md b/changelog.d/1053-governed-integrity.md deleted file mode 100644 index 0699ffc4..00000000 --- a/changelog.d/1053-governed-integrity.md +++ /dev/null @@ -1,6 +0,0 @@ -Changed -`agentdescent governed-harness` accepts a caller-supplied `integrity` pair and -passes it to `govern_harness_change`, so a spec whose after-state reorders the -guardrail path, drops an allow rule, loses a decision record or changes the -action set without declaring it is rejected before proof, approval and -promotion instead of silently skipping the composition check (#1053). diff --git a/changelog.d/1063-span-boundary-redaction.md b/changelog.d/1063-span-boundary-redaction.md deleted file mode 100644 index cca4929f..00000000 --- a/changelog.d/1063-span-boundary-redaction.md +++ /dev/null @@ -1,7 +0,0 @@ -Security -Span values are redacted at the boundary that carries them out of the process: -`set_attribute`, `add_event`, a span's initial `attributes`, a directly -constructed `TelemetrySpan`, a `MemoryOperation.to_span_attributes()` and the -`exception` event's message now pass through the same key/value filter, so a -credential in a precondition, verifier, event detail or raised error is written -as `[redacted]` (#1063). diff --git a/changelog.d/1064-experience-store-leaf.md b/changelog.d/1064-experience-store-leaf.md deleted file mode 100644 index c3ee83b6..00000000 --- a/changelog.d/1064-experience-store-leaf.md +++ /dev/null @@ -1,5 +0,0 @@ -Fixed -`ExperienceStore.add` refuses a link planted at the `--store` leaf before the -first byte — a symlink, a hardlink and a dangling symlink each appended -wherever the link pointed — and refuses a leaf that is not a regular file, -with the sibling writers' `cannot write : ` shape (#1064). diff --git a/changelog.d/1070-secret-key-components.md b/changelog.d/1070-secret-key-components.md deleted file mode 100644 index 623cc286..00000000 --- a/changelog.d/1070-secret-key-components.md +++ /dev/null @@ -1,9 +0,0 @@ -Fixed -Telemetry exports lose no more than credentials: a key is dropped when a whole -component of it names a credential, not when it merely contains the substring -`token` or `secret`. `tokens`, `max_tokens`, `input_tokens`, `output_tokens`, -`token_count`, `tokens_in`, `budget_tokens`, `secret_free` and -`secrets_scanned` are written again (the cost and budget counters among them), -while `api_key`, `apikey`, `access_key`, `client_secret`, `private_key`, -`password`, `authorization`, `credential` and the `*_token` spellings still -drop (#1070). diff --git a/changelog.d/1082-cli-write-dir-enforcement.md b/changelog.d/1082-cli-write-dir-enforcement.md deleted file mode 100644 index 743e7b73..00000000 --- a/changelog.d/1082-cli-write-dir-enforcement.md +++ /dev/null @@ -1,8 +0,0 @@ -Changed -`agentdescent apply` honours `AGENTDESCENT_SURFACE_WRITE_DIRS`: a destination, -a materialised tree or a `.bak-N` slot outside the allowlist is refused with -exit 2 before anything touches disk. Inside a host worker -(`AGENTDESCENT_NESTED=1`) the allowlist defaults to the run's cwd plus -`$AGENTDESCENT_HOME`; outside one it is only enforced when the variable is set, -so a terminal `apply --to ` behaves as before. `show --diff` still -renders and reports the refused target in `apply_plan` (#1082). diff --git a/changelog.d/1082-host-env-scrub.md b/changelog.d/1082-host-env-scrub.md deleted file mode 100644 index 91c88fce..00000000 --- a/changelog.d/1082-host-env-scrub.md +++ /dev/null @@ -1,13 +0,0 @@ -Fixed -A detached run no longer inherits the worker's own session: `runstore.launch` -drops the host-session markers it lists, except the four host config roots it -keeps (`CLAUDE_CONFIG_DIR`, `CODEX_HOME`, `DSH_HOME`, `XDG_CONFIG_HOME`), before -it starts the grandchild, so the host worker's session identity and MCP server -configuration stay with the host while the credentials and `PATH` the run needs -still pass through (#1082). The marker list is `SESSION_MARKERS` (`DSH_*`, -`CLAUDE*`, `CODEX*`, `MCP_*`, the OpenCode config carriers and -`XDG_CONFIG_HOME`), the one the candidate worker already dropped -(`agentdescent.agents.worker_env`); it used to be a second, shorter copy, and -five of its eleven entries crossed the scrub. Those four roots name -configuration rather than session, and the run resolves a host's real home from -them (#1082). diff --git a/changelog.d/1083-otel-wire-path.md b/changelog.d/1083-otel-wire-path.md deleted file mode 100644 index 05a3eca1..00000000 --- a/changelog.d/1083-otel-wire-path.md +++ /dev/null @@ -1,14 +0,0 @@ -Added -An `otel-wire` CI leg installs the `[otel]` extra and exports one span and one -metric to a closed port on `127.0.0.1`, asserting a refused connection is logged -and contained rather than raised out of `export_run`. A test pins that -`export_span` copies `ExportedSpan.attributes` verbatim — it filters and invents -nothing — so the no-content guarantee is `export_run`'s and the caller's for -spans built by hand (#1083). - -The leg also keeps the fail-fast row real where the extra is installed: the SDK -import is halted through the interpreter's own `sys.modules` marker instead of a -probe-and-skip, so construction still has to name the install line, and the -loopback endpoint rows assert the accepted endpoint on a stubbed seam rather than -building a live provider. `docs/metrics.md` now says credentials go in -`OTEL_EXPORTER_OTLP_HEADERS`, never in the URL (#1083). diff --git a/changelog.d/1085-cost-card-fallback.md b/changelog.d/1085-cost-card-fallback.md deleted file mode 100644 index d2e5780c..00000000 --- a/changelog.d/1085-cost-card-fallback.md +++ /dev/null @@ -1,8 +0,0 @@ -Changed -`docs/cost.md` and the `Cost.from_cards` / module docstrings now state what the -card fallback actually returns in-tree: no producer writes -`cost_tokens`/`cost_wallclock` onto an `EvidenceCard`, so `Cost.from_cards` -yields `Cost()` — unknown, not a free diff — and a caller that wants the gate to -price a diff passes `MergeContext.diff_cost = Cost.from_usage(usage)`. The page -also warns that `require_cost=True` defers every candidate on a run whose cards -carry no cost (#1085). diff --git a/changelog.d/1089-noise-refusal-cause.md b/changelog.d/1089-noise-refusal-cause.md deleted file mode 100644 index 3a0ae908..00000000 --- a/changelog.d/1089-noise-refusal-cause.md +++ /dev/null @@ -1,6 +0,0 @@ -Fixed -Trust evidence now names the actual reason a gain is refused as noise: too few -baseline repeats to bound a floor, a gain inside the floor (printed as a band -around the baseline mean rather than an absolute interval against a delta), or -a gain below `min_effect`. `gain_over_noise` is `0.0` when the floor is -undefined instead of a ratio over an epsilon (#1089). diff --git a/changelog.d/1091-negative-gain-note.md b/changelog.d/1091-negative-gain-note.md deleted file mode 100644 index bf7d0b45..00000000 --- a/changelog.d/1091-negative-gain-note.md +++ /dev/null @@ -1,4 +0,0 @@ -Fixed -`trust_evidence` now explains a refused gain that is zero or negative as "the -candidate did not improve on the baseline" instead of blaming the noise floor, -and the note's docstring lists all four causes in code order (#1091). diff --git a/changelog.d/1092-resume-round-numbering.md b/changelog.d/1092-resume-round-numbering.md deleted file mode 100644 index 365f931d..00000000 --- a/changelog.d/1092-resume-round-numbering.md +++ /dev/null @@ -1,7 +0,0 @@ -Fixed -A resumed run continues its ledger's round numbering instead of restarting at -`0`, and `status.best_reward` no longer rewinds below a reward the run's own -history still holds. `execute()` passes one past the last recorded round as the -new `round_base` and seeds the best from the existing rows; a checkpoint further -ahead than the ledger still wins, and an empty ledger still starts at `0` -(#1092). diff --git a/changelog.d/1093-probe-failed-next.md b/changelog.d/1093-probe-failed-next.md deleted file mode 100644 index 45b0edde..00000000 --- a/changelog.d/1093-probe-failed-next.md +++ /dev/null @@ -1,5 +0,0 @@ -Fixed -A failed `plan --probe` or `algorithms run --probe` no longer ends with -`next: agentdescent evolve`. The last line (and the JSON `next`) is the -repair action, `--json` sets `ready: false` while leaving `ok` about the -plan itself, and a successful or omitted probe is unchanged (#1093). diff --git a/changelog.d/1094-mcp-plan-probe.md b/changelog.d/1094-mcp-plan-probe.md deleted file mode 100644 index 9782d077..00000000 --- a/changelog.d/1094-mcp-plan-probe.md +++ /dev/null @@ -1,9 +0,0 @@ -Added -MCP `plan` takes `probe: bool = False` and, when true, runs the same -preflight as `plan --probe`: one minimal call per model-backed agent, with the -rows under `probe` in the result. A refusal makes the plan `ok: false` and -points `next` at the repair action instead of `start`, so a host that follows -`next` no longer discovers an expired subscription only after the run exists. -The default is unchanged: plain `plan` still sends nothing and returns its old -key set. The shared host skill teaches the probe step, so every installed -host now asks for it on the first use of a model (#1094). diff --git a/changelog.d/1095-credential-word-suffix.md b/changelog.d/1095-credential-word-suffix.md deleted file mode 100644 index 892191da..00000000 --- a/changelog.d/1095-credential-word-suffix.md +++ /dev/null @@ -1,9 +0,0 @@ -Fixed -`_is_secret_key` no longer lets a credential through when it arrives as one -word: `PGPASSWORD`, `s3secretkey` and `AWSSecretAccessKey` are matched by -whole-word suffix on a single-component key, and the adjacent-pair lookup -tolerates one plural `s`, so `client_secrets` and `api_keys` are dropped too. -The plural spellings (`passwords`, `pgpasswords`, `db_secrets`) match in the -last position or as a whole component. The `#1070` counters (`tokens`, -`max_tokens`, `secrets_scanned`) still survive, and `auth` stays a recorded -non-goal (#1095). diff --git a/changelog.d/1096-handoff-failed-child.md b/changelog.d/1096-handoff-failed-child.md deleted file mode 100644 index 77ef58e4..00000000 --- a/changelog.d/1096-handoff-failed-child.md +++ /dev/null @@ -1,11 +0,0 @@ -Fixed -The handoff wrapper no longer reports a run that never started as a successful -handoff: when the wait window closes with the child gone and no new -`status.json` where the wrapper can see it, it writes `waiting.json` with -`status: failed` and `state: exited`, and exits nonzero with the child's own -reason left in `worker.log`. A child that is still running and has not written -its status yet is unchanged: `waiting`, `found: false`, exit 0. A store the -caller owns as the run directory is still reported from its own rewritten -`status.json`. The verdict reads the store again after the liveness probe, so a -status written while the probe ran is still this handoff's. The page says all -three (#1096). diff --git a/changelog.d/1098-env-scrub-exception.md b/changelog.d/1098-env-scrub-exception.md deleted file mode 100644 index 166166c8..00000000 --- a/changelog.d/1098-env-scrub-exception.md +++ /dev/null @@ -1,6 +0,0 @@ -Fixed -The env-scrub note no longer says it drops "every" session marker and then keeps -some of them: `docs/dsh-super-agent.md` and the #1082 release note now name the -four host config roots (`CLAUDE_CONFIG_DIR`, `CODEX_HOME`, `DSH_HOME`, -`XDG_CONFIG_HOME`) as an exception to the marker list before that list appears, -so the first clause cannot be read as dropping them (#1098). diff --git a/changelog.d/1100-manifest-secret-keys.md b/changelog.d/1100-manifest-secret-keys.md deleted file mode 100644 index 28b24008..00000000 --- a/changelog.d/1100-manifest-secret-keys.md +++ /dev/null @@ -1,10 +0,0 @@ -Fixed -`manifest.redact` now matches every credential spelling `rsi_telemetry` does: -`PGPASSWORD`, `s3secretkey`, `client_secrets` and the other one-word or plural -forms are redacted before hashing and before a manifest or trajectory record is -written, instead of being stored and hashed in clear. The manifest predicate -keeps its own rules (`bearer`, and the `*_key` suffix that `sort_key` relies on) -and delegates the rest, so a key redacted before still is. Records whose -`model_config` or `extra` carried one of the newly matched spellings get a new -`model_hash` / `spec_hash`, so `verify_manifest` reports those inputs as changed -for them: that is the fix redacting what it used to leak, not a bug (#1100). diff --git a/changelog.d/1101-handoff-run-epoch.md b/changelog.d/1101-handoff-run-epoch.md deleted file mode 100644 index 38633bc3..00000000 --- a/changelog.d/1101-handoff-run-epoch.md +++ /dev/null @@ -1,7 +0,0 @@ -Fixed -The `scripts/dsh-superagent-run.sh` handoff no longer reports an earlier run that -is still going. The launch epoch is taken with the `status.json` snapshot, and a -changed file is only reported when its `started` is numeric and no older than -that epoch, so a handoff into a reused store names its own run — or -`found: false` — instead of the earlier run, whose `run_id` the page's read-back -would then have polled (#1101). diff --git a/changelog.d/1106-algorithms-run-hint.md b/changelog.d/1106-algorithms-run-hint.md deleted file mode 100644 index 882a387f..00000000 --- a/changelog.d/1106-algorithms-run-hint.md +++ /dev/null @@ -1,7 +0,0 @@ -Fixed -`algorithms run` no longer prints the `--rounds 1 --workers 1` quick-trial hint -after a refused `--probe`: with the kind's default rounds and workers the hint -used to land below `next: renew the subscription …`, so the last line on screen -answered a question the refusal had already answered. The hint still prints -after a passing probe and when no probe runs, and every other line of the -refusal is unchanged (#1106). diff --git a/changelog.d/1107-demo-tool-description-evoskill.md b/changelog.d/1107-demo-tool-description-evoskill.md deleted file mode 100644 index ad0c0ca2..00000000 --- a/changelog.d/1107-demo-tool-description-evoskill.md +++ /dev/null @@ -1,5 +0,0 @@ -Fixed -The MCP `demo` tool description no longer says evoskill needs a real model: -it lists the same offline cards as `agentdescent demo --help` (including -evoskill, which defaults to the `skill_library` fixture) and accepts -`kind` `skill_library` (#1107). diff --git a/changelog.d/1108-algorithm-spec-default-out.md b/changelog.d/1108-algorithm-spec-default-out.md deleted file mode 100644 index 2a00cfe6..00000000 --- a/changelog.d/1108-algorithm-spec-default-out.md +++ /dev/null @@ -1,7 +0,0 @@ -Fixed -`algorithm_spec` no longer defaults its starter spec to a path beside the -user's target, which the default write allowlist refused: an omitted `out` now -writes `agentdescent..spec.json` under `$AGENTDESCENT_HOME` (or the first -explicitly allowed root), so the shortest documented path plans on an untouched -install. The tool description and the shared `SKILL.md` say so, and a refused -explicit `out` names the allowed roots and an `out` that works (#1108). diff --git a/changelog.d/1109-store-quote.md b/changelog.d/1109-store-quote.md deleted file mode 100644 index bdd6ca0b..00000000 --- a/changelog.d/1109-store-quote.md +++ /dev/null @@ -1,6 +0,0 @@ -Security -The handoff wrapper no longer builds its launch command as a shell string: the -log path and the run's command reach `tmux`/`setsid` as arguments, so a `--store` -containing `'`, `;`, `$` or a space is used literally instead of closing the -quote and running the rest of the path as shell words outside the worker's -process group (#1109). diff --git a/changelog.d/1112-lab-grouping.md b/changelog.d/1112-lab-grouping.md deleted file mode 100644 index 2cec8312..00000000 --- a/changelog.d/1112-lab-grouping.md +++ /dev/null @@ -1,6 +0,0 @@ -Changed -`agentdescent --help` now lists the first-run path (`demo`, `init`, `plan`, -`evolve`, `show`, `apply`, `doctor`, `status`) plus `mcp`, and every other verb -is grouped under a new `lab` subcommand (`agentdescent lab campaign …`); each -verb keeps its old top-level path, its arguments, its output and its exit code, -and the same declaration mounts both paths (#1112). diff --git a/changelog.d/1113-plan-headroom-artifact-shape.md b/changelog.d/1113-plan-headroom-artifact-shape.md deleted file mode 100644 index a9264f16..00000000 --- a/changelog.d/1113-plan-headroom-artifact-shape.md +++ /dev/null @@ -1,10 +0,0 @@ -Added -`plan` can now catch two wasteful runs before they are paid for. The opt-in -`--baseline-sample N` scores the seed artifact on the first N tasks (one agent -call per task, quoted before the first call) and a seed at or above -`--saturation` (default `0.90`) prints `no headroom` and withholds the paid -run's quote unless `--force` is passed; `--json` carries the result as -`baseline` and sets `ready: false`. A plain `plan` still sends nothing. -`plan` and the pilot report also say when the artifact holds one key (a -`SingleSlot` prompt, a one-file tree): worker proposals contradict on that key, -so extra workers buy best-of-N selection, not merging (#1113). diff --git a/changelog.d/1118-apply-created-target.md b/changelog.d/1118-apply-created-target.md deleted file mode 100644 index ec1cb259..00000000 --- a/changelog.d/1118-apply-created-target.md +++ /dev/null @@ -1,10 +0,0 @@ -Fixed -`apply` into a tree target that no longer exists now says so: the plan carries -`"created": true` and a real run prints `created from run ; nothing -was restored and no backup was taken` once on stderr, instead of a -`written`/`backup: []` plan that read like an ordinary in-place update. An -in-place apply keeps `"created": false` and prints no line; so does a no-op one, -because an empty artifact into a missing target writes nothing and must not -claim a recreate. The dry run reports the flag without the line, since it -creates nothing. The flag is emitted for the tree path only: a text `--to` plan -is unchanged and carries no `created` key (#1118). diff --git a/changelog.d/1119-wrapped-trace-digest.md b/changelog.d/1119-wrapped-trace-digest.md deleted file mode 100644 index 660f1db7..00000000 --- a/changelog.d/1119-wrapped-trace-digest.md +++ /dev/null @@ -1,4 +0,0 @@ -Fixed -`traces digest --events FILE` now reads a wrapped `{"events": [...]}` object -(compact or pretty-printed) as the same digest as a JSON array, instead of -reporting a clean bill or exiting 2 (#1119). diff --git a/changelog.d/1120-publisher-fast-forward-docs.md b/changelog.d/1120-publisher-fast-forward-docs.md deleted file mode 100644 index 827cba66..00000000 --- a/changelog.d/1120-publisher-fast-forward-docs.md +++ /dev/null @@ -1,4 +0,0 @@ -Fixed -The GitHub publisher page now documents fast-forward-only proposal branch -updates: a second publish of a `behind` or `diverged` branch is refused with -`refusing to rewrite it` instead of promising a force update (#1120). diff --git a/changelog.d/1122-algorithm-spec-target-dir.md b/changelog.d/1122-algorithm-spec-target-dir.md deleted file mode 100644 index b1ab197b..00000000 --- a/changelog.d/1122-algorithm-spec-target-dir.md +++ /dev/null @@ -1,8 +0,0 @@ -Fixed -`algorithm_spec` no longer writes its default starter spec inside a directory -target: when the only allowed write roots are the target itself or below it, -the call returns `ok: false` naming the target and -AGENTDESCENT_SURFACE_WRITE_DIRS instead of -dropping `agentdescent..spec.json` into the user's directory; a usable -root outside the target is still preferred, and the description no longer -claims `$AGENTDESCENT_HOME` when that home is the target (#1122). diff --git a/changelog.d/1124-text-apply-missing-parent.md b/changelog.d/1124-text-apply-missing-parent.md deleted file mode 100644 index fe293d88..00000000 --- a/changelog.d/1124-text-apply-missing-parent.md +++ /dev/null @@ -1,5 +0,0 @@ -Fixed -A text-kind `apply --to /rule.md` now creates the missing parent -directory before writing, so the dry run it had already reported green is true: -the real run exits 0 and the file lands at that path. A genuinely unwritable -`--to` still exits 2 with the single `error: cannot write …` line (#1124). diff --git a/changelog.d/1130-star-history.md b/changelog.d/1130-star-history.md deleted file mode 100644 index a5eb0903..00000000 --- a/changelog.d/1130-star-history.md +++ /dev/null @@ -1,3 +0,0 @@ -Added -A Star History chart at the end of the README, so visitors can see the -repository's GitHub star growth over time (#1130). diff --git a/changelog.d/1131-docs-first-run-nav.md b/changelog.d/1131-docs-first-run-nav.md deleted file mode 100644 index d87b02e5..00000000 --- a/changelog.d/1131-docs-first-run-nav.md +++ /dev/null @@ -1,4 +0,0 @@ -Changed -The README and documentation landing pages now lead with one first-run path, -and the docs navigation is grouped by getting started, hosts and plugins, and -algorithms, instead of one long mixed list (#1131). diff --git a/changelog.d/1132-cost-page-pilot-shipped.md b/changelog.d/1132-cost-page-pilot-shipped.md deleted file mode 100644 index fd27adb1..00000000 --- a/changelog.d/1132-cost-page-pilot-shipped.md +++ /dev/null @@ -1,4 +0,0 @@ -Fixed -The cost page no longer says `agentdescent/pilot.py` is absent: the module shipped -in #1121, its `PilotReport` carries the run's cost mapping, and `CostReport` -stays the cost-only surface that page describes (#1132). diff --git a/changelog.d/1137-pilot-multi-seed-artifact.md b/changelog.d/1137-pilot-multi-seed-artifact.md deleted file mode 100644 index 0d4eda88..00000000 --- a/changelog.d/1137-pilot-multi-seed-artifact.md +++ /dev/null @@ -1,4 +0,0 @@ -Fixed -`pilot` no longer reports the seed-0 artifact next to scores pooled -across every seed. A single artifact is reported only when every seed -produced it; otherwise the report summarizes per seed (#1137). diff --git a/changelog.d/1138-artifact-snapshot.md b/changelog.d/1138-artifact-snapshot.md deleted file mode 100644 index f1ac855f..00000000 --- a/changelog.d/1138-artifact-snapshot.md +++ /dev/null @@ -1,6 +0,0 @@ -Added -Portable artifact snapshots capture a complete FileTree candidate -(editable files plus frozen context) and the ledger's existing -id/parent/root lineage, and can be re-scored on another compatible -dataset without further search (`agentdescent snapshot -export|show|eval|apply`, #1138). diff --git a/changelog.d/1141-pilot-next-artifact.md b/changelog.d/1141-pilot-next-artifact.md deleted file mode 100644 index 73b15173..00000000 --- a/changelog.d/1141-pilot-next-artifact.md +++ /dev/null @@ -1,6 +0,0 @@ -Fixed -A promoting multi-seed pilot report no longer tells the reader to "review the -artifact diff" or "apply it" when the seeds produced different artifacts and the -report names none. The steps point at `run.per_seed` and ask for one seed's -artifact to be chosen, reviewed and applied as an explicit step; the single-seed -and agreeing-seed reports keep their current text (#1141). diff --git a/changelog.d/1145-demo-algorithm-next-evolve.md b/changelog.d/1145-demo-algorithm-next-evolve.md deleted file mode 100644 index b18bda2b..00000000 --- a/changelog.d/1145-demo-algorithm-next-evolve.md +++ /dev/null @@ -1,6 +0,0 @@ -Fixed -`demo --algorithm NAME` no longer ends with `evolve agentdescent.NAME.spec.json`. -That file does not exist until `algorithms run` writes it, so following the -footer (or `--json` / MCP `next`) failed with exit 2. The footer still points -at `algorithms run NAME ...`, which plans and then prints `evolve` with the -absolute path it just wrote (#1145). diff --git a/changelog.d/1146-first-run-transcript.md b/changelog.d/1146-first-run-transcript.md deleted file mode 100644 index 6534df00..00000000 --- a/changelog.d/1146-first-run-transcript.md +++ /dev/null @@ -1,5 +0,0 @@ -Fixed -The first-screen demo transcripts (`README.md`, `docs/start.md`, -`docs/plugin-quickstart.md`, `docs/testing-guide.md`) now show the seed baseline -the CLI prints, `held-out reward: 0.000 -> 1.000`, instead of a bare `1.000`, and -a docs guard keeps those four blocks on the command's output (#1146). diff --git a/changelog.d/1149-snapshot-apply-text.md b/changelog.d/1149-snapshot-apply-text.md deleted file mode 100644 index be9fc08e..00000000 --- a/changelog.d/1149-snapshot-apply-text.md +++ /dev/null @@ -1,2 +0,0 @@ -Fixed -`snapshot apply` on a text snapshot now writes the rendered prompt to the file named by `--to`, with the same backup, no-op, and parent-creation rules as `apply --from`. It previously materialised the strategy's internal rule store (hash-named files) into a directory, and an empty-rules snapshot reported a no-op instead of writing the seed prompt. A directory `--to` is refused. Tree-kind snapshots are unchanged (#1149). diff --git a/changelog.d/1150-snapshot-artifact-id.md b/changelog.d/1150-snapshot-artifact-id.md deleted file mode 100644 index bc1c966b..00000000 --- a/changelog.d/1150-snapshot-artifact-id.md +++ /dev/null @@ -1,8 +0,0 @@ -Fixed -A snapshot's `artifact_id` is validated as one path-safe component when the -snapshot is built, and both staging writes (`snapshot eval`, `snapshot apply`) -stage under a fixed name inside their temp directory instead of deriving a -path from the id: applying or scoring a snapshot JSON received from someone -else can no longer make the staging write land outside that temp directory. -A malformed id is reported as a clean CLI error (exit 2), not a traceback -(#1150). diff --git a/changelog.d/1154-snapshot-eval-spec-json.md b/changelog.d/1154-snapshot-eval-spec-json.md deleted file mode 100644 index 7a6f6300..00000000 --- a/changelog.d/1154-snapshot-eval-spec-json.md +++ /dev/null @@ -1,3 +0,0 @@ -Fixed -The snapshot re-evaluation examples now use `dest.json`, which `load_spec` -can read, instead of `dest.yaml` (#1154). diff --git a/changelog.d/1155-baseline-sample-failed-envelope.md b/changelog.d/1155-baseline-sample-failed-envelope.md deleted file mode 100644 index bcc77d6a..00000000 --- a/changelog.d/1155-baseline-sample-failed-envelope.md +++ /dev/null @@ -1,7 +0,0 @@ -Fixed -`plan --baseline-sample` now keeps the plan envelope when seed scoring fails -on a provider, subscription, or endpoint error: exit 1 (not 2), `--json` -writes `ok: true`, `ready: false`, a `baseline` object with the failure, and -a `next` repair step (the same wording `plan --probe` already uses), and the -human form keeps the plan plus `next:` / `then:`. Input errors such as -`--baseline-sample 0` still exit 2. No run-store entry is written (#1155). diff --git a/changelog.d/1157-baseline-sample-quote-actual-tasks.md b/changelog.d/1157-baseline-sample-quote-actual-tasks.md deleted file mode 100644 index 8c6942be..00000000 --- a/changelog.d/1157-baseline-sample-quote-actual-tasks.md +++ /dev/null @@ -1,5 +0,0 @@ -Fixed -`plan --baseline-sample` now quotes the number of tasks the sample can -actually score, not the requested N: a 4-task spec asked for 50 is priced -at 4 calls, and that count matches `baseline.sample`. A request below the -task count is unchanged (#1157). diff --git a/changelog.d/1161-claim-expiry-refs.md b/changelog.d/1161-claim-expiry-refs.md deleted file mode 100644 index 7293d4df..00000000 --- a/changelog.d/1161-claim-expiry-refs.md +++ /dev/null @@ -1,5 +0,0 @@ -Fixed -A claim no longer expires under a pull request that links its issue with -`Refs #N` or `Part of #N`: the controller now reads those references when it -decides whether the claim's TTL is still fed by a push, while only closing -keywords still make a PR an agent PR (#1161). diff --git a/changelog.d/1164-snapshot-eval-skill-name.md b/changelog.d/1164-snapshot-eval-skill-name.md deleted file mode 100644 index 8c79b68b..00000000 --- a/changelog.d/1164-snapshot-eval-skill-name.md +++ /dev/null @@ -1,6 +0,0 @@ -Fixed -`snapshot eval` of a `skill_dir` snapshot now exposes the candidate under -the snapshot's `artifact_id` in the destination layout (for example -`.claude/skills/csv-total/`), so a demo export re-evaluated with `--spec` -scores the learned skill instead of 0 because the dest agent looked under -`.claude/skills/artifact/` (#1164). diff --git a/changelog.d/1166-run-call-cap.md b/changelog.d/1166-run-call-cap.md deleted file mode 100644 index 01b6a04b..00000000 --- a/changelog.d/1166-run-call-cap.md +++ /dev/null @@ -1,11 +0,0 @@ -Added -A run-level call cap that holds before the call (`--max-calls`, `--max-usd`, -also `max_calls` / `max_usd` in a spec's `evolve` block) (#1166). Unlike -`max_calls` on `evolve()`, which is a round barrier and lets a four-worker round -overshoot, the guarded cap is charged before each model call: the call that -would cross the limit is not made, the round closes without a partial merge or a -traceback, and the run ends `done` with `stop_reason=budget` and a -`stop_detail` naming the limit and the observed value, in `status.json` and in -`agentdescent status`. `0` is a real limit (no calls); a negative or non-numeric -value is refused at validation with the field and the value named; with neither -flag set the run and `status.json` are unchanged. Offline, no new dependency. diff --git a/changelog.d/1171-lab-hides-internal-run.md b/changelog.d/1171-lab-hides-internal-run.md deleted file mode 100644 index eb6f0c91..00000000 --- a/changelog.d/1171-lab-hides-internal-run.md +++ /dev/null @@ -1,4 +0,0 @@ -Fixed -`agentdescent lab --help` and the usage line on a `lab` argparse error no -longer list the internal `run` worker in the `{...}` metavar; the verb stays -reachable as `lab run` (#1171). diff --git a/changelog.d/1174-help-epilog.md b/changelog.d/1174-help-epilog.md deleted file mode 100644 index 1d2e70c8..00000000 --- a/changelog.d/1174-help-epilog.md +++ /dev/null @@ -1,7 +0,0 @@ -Changed -The top-level `agentdescent --help` now ends with a short next-steps block -naming `agentdescent install `, `agentdescent algorithms list` and -`agentdescent lab --help`, so the two verbs the first-run docs use are -discoverable on the first screen without reading the ~40 verbs under `lab` -(#1174). The listing, the usage line and `DEFAULT_SURFACE` are unchanged from -#1112, and no verb, argument, output or exit code changed. diff --git a/changelog.d/912-mcp-http-bearer-token.md b/changelog.d/912-mcp-http-bearer-token.md deleted file mode 100644 index c53af31e..00000000 --- a/changelog.d/912-mcp-http-bearer-token.md +++ /dev/null @@ -1,6 +0,0 @@ -Security -`agentdescent mcp --transport streamable-http` can now require an -`Authorization: Bearer` header (`--http-token-env NAME`, with the token read -from the environment and never from the command line). A request without it is -answered 401 before an MCP session is created, so a leaked tunnel URL no longer -reaches `apply` or `start`; `--allow-remote` is refused without a token (#912). diff --git a/changelog.d/920-embedding-facade.md b/changelog.d/920-embedding-facade.md deleted file mode 100644 index 04d16f4a..00000000 --- a/changelog.d/920-embedding-facade.md +++ /dev/null @@ -1,8 +0,0 @@ -Added -`agentdescent.facade` is the versioned import surface for embedding the engine: -a curated `__all__` covering Path A and the durable `LoopKernel` path, with -`API_VERSION = "1"` and a promise that changes are additive within a version and -removals need a bump. `tests/test_facade.py` pins the list, so widening or -narrowing it is a deliberate edit rather than an accidental import; nothing -outside the list — including the package root's research exports — is covered -(#920). diff --git a/changelog.d/923-histogram-method.md b/changelog.d/923-histogram-method.md deleted file mode 100644 index 525e80f0..00000000 --- a/changelog.d/923-histogram-method.md +++ /dev/null @@ -1,6 +0,0 @@ -Fixed -`export_metric()` sent histograms through the counter's `add`, so the first real -export of an `agentdescent_rsi_round_reward` raised `AttributeError` out of -`export_run`. Histograms now call the SDK's `Histogram.record()`, and the fake -SDK gives each instrument kind only its own method, so the same mistake fails in -the offline suite instead of on a host's first run (#923). diff --git a/changelog.d/923-otlp-exporter.md b/changelog.d/923-otlp-exporter.md deleted file mode 100644 index 9cb6fa69..00000000 --- a/changelog.d/923-otlp-exporter.md +++ /dev/null @@ -1,12 +0,0 @@ -Added -An opt-in OpenTelemetry bridge (`agentdescent.otel_export`): `export_run()` writes -one span per round and one per decision, and re-publishes the `rsi_metrics` -families plus a round-reward histogram. It is off unless -`OTEL_EXPORTER_OTLP_ENDPOINT` is set, a host can inject any object with -`export_span`/`export_metric` (so tests need no collector), and the new optional -`agentdescent[otel]` extra is imported only when an `OtlpExporter` is -constructed — importing the package and running normally needs nothing new. No -prompts or responses are exported by default. An endpoint carrying -credentials is refused without the value being echoed, and plain -`http://` is accepted only for a literal loopback collector; a remote -collector needs TLS terminated in front of it (#923). diff --git a/changelog.d/924-artifact-store-protocol.md b/changelog.d/924-artifact-store-protocol.md deleted file mode 100644 index 07c6a3d5..00000000 --- a/changelog.d/924-artifact-store-protocol.md +++ /dev/null @@ -1,4 +0,0 @@ -Added -`agentdescent.ports` adds the `ArtifactStore` protocol and the -`FileArtifactStore` reference backend over a content-addressed directory tree; -an id, revision, parent or path that could leave the root is refused (#1026). diff --git a/changelog.d/924-evaluation-provider-protocol.md b/changelog.d/924-evaluation-provider-protocol.md deleted file mode 100644 index 58abc7cf..00000000 --- a/changelog.d/924-evaluation-provider-protocol.md +++ /dev/null @@ -1,6 +0,0 @@ -Added -`agentdescent.ports` adds the `EvaluationProvider` protocol and the -`CsvEvaluationProvider` reference backend, so a split and the scorer a run -optimises can come from a database or an internal judge service instead of a -file; `oracle` and `judge` report `None` rather than a score they cannot measure -(#924). diff --git a/changelog.d/928-superseding-report.md b/changelog.d/928-superseding-report.md deleted file mode 100644 index 0a2e4812..00000000 --- a/changelog.d/928-superseding-report.md +++ /dev/null @@ -1,7 +0,0 @@ -Fixed -The review gate now orders comments and reviews together by time and lets the -newest report at a tip decide, so a same-tip follow-up with no `Verdict:` line -(a withdrawal, an added requirement, an open question) voids an earlier PASS -there instead of leaving it counting. A report that names another tip is still -judged against that tip, a newer PASS still counts, and the `Ran:` / `Not run:` -handling for the winning PASS is unchanged (#928). diff --git a/changelog.d/937-pilot-core.md b/changelog.d/937-pilot-core.md deleted file mode 100644 index 30e5c3a1..00000000 --- a/changelog.d/937-pilot-core.md +++ /dev/null @@ -1,10 +0,0 @@ -Added -`agentdescent.pilot` turns one evolution run into one report a user can defend: -the seed and the result are scored on the same disjoint test tasks, the paired -scores become `TrustEvidence`, and `promote()` returns `promote` / `defer` / -`reject` with the reasons, cost, manifest and next action. A test task that -raised is run-health evidence rather than a zero, an unchanged artifact reports -`noise_only`, a missing contamination check defers, and a contaminated dataset -rejects before any model call. `pilot_scores()` exposes the same decision seam -for a run that already happened elsewhere; the CLI and `apply` are separate -follow-ups (#937). diff --git a/changelog.d/938-falsification-acceptance.md b/changelog.d/938-falsification-acceptance.md deleted file mode 100644 index e1561f29..00000000 --- a/changelog.d/938-falsification-acceptance.md +++ /dev/null @@ -1,16 +0,0 @@ -Added -Falsification-gated lesson admission: `FalsificationAcceptance` fills the existing -`Policies(acceptance=...)` slot and decides a commit on per-task measurements -instead of held-out counts alone. A lesson commits only with a measured gain on the -target-transfer partition, no per-task regression on the archived partition, and no -verdict change on the falsification partition; each reject reason -(`no-target-transfer`, `archived-regression`, `falsification-moved`) names the -failing task ids. Probes are caller-supplied (`MergeContext.probes`), and when they -are unavailable the policy defers (`AcceptDecision(defer=True)`) rather than falling -back to the count-only gate; all three partitions are required, and an empty one is -refused at construction instead of deferring on every candidate. A deferral under a -category this build does not know reports `probes-unavailable` rather than raising -inside the merge thread. A committed lesson hands its review evidence -(`rule_id -> {precondition, behavior_change, expected_effect, verifier, trace_ids, -evidence_card_hash}`) to the shared provenance writer as -`AcceptDecision.provenance` (#938). diff --git a/changelog.d/940-verifier-admission-universality.md b/changelog.d/940-verifier-admission-universality.md deleted file mode 100644 index 38748344..00000000 --- a/changelog.d/940-verifier-admission-universality.md +++ /dev/null @@ -1,7 +0,0 @@ -Added -`VerifierAgreement.admission_report()` now measures the cross-prompt error -overlap of the accepted stratum: the share of audit prompts on which a commit -the verifier accepted is oracle-denied, reported as `repeat_overlap`, with -`high_universality` at half the prompts over three audited accepts. A high -overlap blocks the admit the mean gains would allow; an unmeasurable one (a -single prompt, or a thin sample) stays `None` and never reads as zero (#940). diff --git a/changelog.d/940-verifier-admission.md b/changelog.d/940-verifier-admission.md deleted file mode 100644 index 01ce1702..00000000 --- a/changelog.d/940-verifier-admission.md +++ /dev/null @@ -1,11 +0,0 @@ -Added -Verifier admission on both strata: `VerifierAgreement(rejected_sample_rate=...)` -audits a sample of the commits it rejected, and `admission_report()` reports -precision on the accepted stratum with its bootstrap interval and -`rejection_accuracy_{lower,upper}`, the interval on the share of rejections that -were correct, so a verifier that rejects good work no longer scores 1.0 on -precision alone. A demote needs a bound already under 0.5 on a stratum with at -least three audited decisions, an admit needs both pessimistic bounds at 0.8, and -a thin sample defers. `require_verifier_admission=True` carries it on -`EnsembleReport.admission`: missing, incomplete or insufficient defers, `reject` -fails closed (#940). diff --git a/changelog.d/944-noise-calibration.md b/changelog.d/944-noise-calibration.md deleted file mode 100644 index 9359761f..00000000 --- a/changelog.d/944-noise-calibration.md +++ /dev/null @@ -1,5 +0,0 @@ -Added -Baseline noise calibration repeats the baseline and reports a `noise_floor`, -task/model/run variance decomposition, `gain_over_noise` and `distinguishable`; -a gain inside the floor is labelled "not distinguishable from noise" and is -never promotion evidence (#944). diff --git a/changelog.d/945-cost-per-accepted-diff.md b/changelog.d/945-cost-per-accepted-diff.md deleted file mode 100644 index 5c79c128..00000000 --- a/changelog.d/945-cost-per-accepted-diff.md +++ /dev/null @@ -1,13 +0,0 @@ -Added -A per-diff `Cost` block (calls, tokens, wall seconds, optional USD), the -`cost_efficiency` / `cost_status` reads, a `CostPolicy` acceptance wrapper that -refuses a quality gain whose cost efficiency regressed (and defers when cost is -required but unmeasured), a `cost_report` with cost per accepted diff and per -point of gain, and `PromotionRules(require_cost=True)` (#945). A report average -is `None` unless every decision measured the unit being divided -- a row priced -in tokens only cannot make the dollar figure look cheaper than the evidence. -This also fixes the API reference for the name it exports: `docs/api.md` was -generated from a flat short-name map, so `agentdescent.Cost` was shadowed by -`agentdescent.audit.scorecard.Cost` and the page documented a signature the -top-level import does not return. The generator now keys by `(module, name)` and -`agentdescent.cost` has its own section, so both types are documented. diff --git a/changelog.d/946-failure-stratified-sampler.md b/changelog.d/946-failure-stratified-sampler.md deleted file mode 100644 index f174126c..00000000 --- a/changelog.d/946-failure-stratified-sampler.md +++ /dev/null @@ -1,7 +0,0 @@ -Added -`FailureStratifiedSampler` stratifies tasks by base failure signature, keeps -search/select/final sets disjoint, and reports stratum coverage, overlap and -transfer status so one dominant failure mode cannot take over the search set -(#946). It is exported from the package root, the transfer comparison keeps a -float floor so equal means report `holds`, and the pick refusal names the -supported wiring (`tasks=sampler.split().search`). diff --git a/changelog.d/947-memory-spans.md b/changelog.d/947-memory-spans.md deleted file mode 100644 index e9cd9400..00000000 --- a/changelog.d/947-memory-spans.md +++ /dev/null @@ -1,7 +0,0 @@ -Added -Memory operation telemetry: `memory.read` / `memory.write` / `memory.update` / -`memory.forget` spans with lesson provenance (precondition, evidence hash, -source trace ids, verifier, expected effect), plus a memory audit block on the -telemetry and experience reports that counts each operation, flags lessons -without provenance, and lists forgotten or rolled-back lessons with a reason and -attributable regressions (#947). diff --git a/changelog.d/949-harness-integrity.md b/changelog.d/949-harness-integrity.md deleted file mode 100644 index 8e7ccdac..00000000 --- a/changelog.d/949-harness-integrity.md +++ /dev/null @@ -1,8 +0,0 @@ -Added -`HarnessIntegrity` checks a plugin/harness before/after snapshot before -promotion: guardrail enforcement-point order (policy before tool mediation), -permission and allow-rule state, the action set unless the diff declares exactly -that change, and the decision-path record for every enforced relationship. -`govern_harness_change(..., integrity=...)` rejects a broken composition before -the proof, approval and promotion seams, and the report names each broken -relationship with the state it observed (#949). diff --git a/changelog.d/951-evidence-card-1042.md b/changelog.d/951-evidence-card-1042.md deleted file mode 100644 index adf1c178..00000000 --- a/changelog.d/951-evidence-card-1042.md +++ /dev/null @@ -1,6 +0,0 @@ -Added -`reports/evidence/1042-ports-leaf-write-through.json` publishes the -leaf-write-through fix as a v1 evidence card — `code` and `reliability` cite -the merged sha, the guarded lines and the regression test, and the other ten -fields say `not reviewed` — with a test that keeps every published card -schema-valid (#951). diff --git a/changelog.d/951-evidence-card-check.md b/changelog.d/951-evidence-card-check.md deleted file mode 100644 index 85e5153a..00000000 --- a/changelog.d/951-evidence-card-check.md +++ /dev/null @@ -1,5 +0,0 @@ -Added -`python -m tools.evidence_card --check` validates the 12-field evidence card -schema v1 over the named `*.json` cards, and refuses a missing, unknown or blank -field by name; `not reviewed` is the explicit way to say a field has no evidence -(#1025). diff --git a/changelog.d/951-evidence-cards-headline.md b/changelog.d/951-evidence-cards-headline.md deleted file mode 100644 index 08f04bc3..00000000 --- a/changelog.d/951-evidence-cards-headline.md +++ /dev/null @@ -1,6 +0,0 @@ -Added -Two evidence cards for headline results from `reports/` -- the audit's constant -rubric tying the judge at 0.836, and the 177-pair verifier bias with its 0.2203 -resolution floor -- plus `reports/evidence/README.md`, which states that a new -claim needs a card, and tests pinning each card's number, its reproduction and -the rule itself (#951). diff --git a/changelog.d/951-evidence-cards.md b/changelog.d/951-evidence-cards.md deleted file mode 100644 index 68c48bfb..00000000 --- a/changelog.d/951-evidence-cards.md +++ /dev/null @@ -1,11 +0,0 @@ -Changed -The results and audit pages now point at evidence cards for their headline -claims: one card per claim under `reports/evidence/`, each answering all twelve -fields and writing `not reviewed` where nobody measured. Eight new cards cover -the one-call path, bringing your own agent, the efficiency rows, equal-budget -HotpotQA, the merge cost lever, judge bias across five workloads, ground-truth -repair, and the rule search's residual against the sigma floor. The directory -indexes them in `reports/evidence/README.md`, the two bench cards are recomputed -in CI from the committed runs, and the rest say on the card that their number -came from a live run — so the gap between "we measured it" and "you can check -it" is visible where the claim is (#951). diff --git a/changelog.d/962-dsh-super-agent.md b/changelog.d/962-dsh-super-agent.md deleted file mode 100644 index 45d96f55..00000000 --- a/changelog.d/962-dsh-super-agent.md +++ /dev/null @@ -1,6 +0,0 @@ -Added -`docs/dsh-super-agent.md` and `scripts/verify-dsh-superagent.sh`: the zero-code -composition of AgentDescent with dsh-super-agent — the two Cordis rows in the -`superagent` profile, the skill-or-MCP fallback, the `plan → start → status → -show` worker loop with no auto-apply, write-dir and budget settings for a -worker, and a gated check that skips cleanly without a live host (#962). diff --git a/changelog.d/963-nested-run-refusal.md b/changelog.d/963-nested-run-refusal.md deleted file mode 100644 index 575fd561..00000000 --- a/changelog.d/963-nested-run-refusal.md +++ /dev/null @@ -1,7 +0,0 @@ -Added -A worker session (`AGENTDESCENT_NESTED=1`, set in every host worker the runner -spawns) can no longer start a nested run through the CLI: `agentdescent evolve` -refuses with exit 2 and `reason: nested-run` before it reads the spec or creates -a run, and `doctor` reports `nested_guard` instead of only a boolean. The MCP -`start` stub carries the same reason and sentence, so a host branches on one -contract. Offline, no new sandbox or auth layer (#963). diff --git a/changelog.d/964-long-run-handoff.md b/changelog.d/964-long-run-handoff.md deleted file mode 100644 index 7c667d31..00000000 --- a/changelog.d/964-long-run-handoff.md +++ /dev/null @@ -1,6 +0,0 @@ -Added -`scripts/dsh-superagent-run.sh` hands a long evolution run off a -dsh-superagent worker: it starts the run in a detached `tmux` session -(`setsid` fallback), writes `waiting.json` with the run id, round and run -directory, and returns at once, so a worker turn no longer holds its slot for -the length of the run (#964). diff --git a/docs/assets/experience.js b/docs/assets/experience.js index 57a60caf..52106312 100644 --- a/docs/assets/experience.js +++ b/docs/assets/experience.js @@ -434,7 +434,7 @@ return `"${cli}": ${demoHosts[k] === host ? `"/usr/local/bin/${cli}"` : 'null'}`; }).join(', '); if (!await toolCall('doctor', '', [ - '{"version": "0.10.0",', + '{"version": "0.11.0",', ` "agent_clis": {${clis}, "git": "/usr/bin/git"},`, ' "provider_keys_present": {"ANTHROPIC_API_KEY": false, "OPENAI_API_KEY": false, "ARK_API_KEY": true},', ` "run_store": "${HOME}/.agentdescent/runs",`, diff --git a/docs/docs.md b/docs/docs.md index c7e571bc..cc7fc0ee 100644 --- a/docs/docs.md +++ b/docs/docs.md @@ -27,7 +27,7 @@ deeper reference when you need it. ### Getting started 1. **Install and see the loop once, offline.** - `pip install "agentdescent[mcp]>=0.10.0"`, then `agentdescent demo`. + `pip install "agentdescent[mcp]>=0.11.0"`, then `agentdescent demo`. Skill evolution means a multi-file library: `agentdescent demo --kind skill_library`. [Install and first run](start.md) diff --git a/docs/index.md b/docs/index.md index 0cbb11b5..9d2e1210 100644 --- a/docs/index.md +++ b/docs/index.md @@ -12,7 +12,7 @@ description: AgentDescent is a modular evolve engine for prompts, skills, progra
-
OPEN SOURCE · v0.10.0 · BUILT FOR CODING AGENTS

Evolve
anything.

Diffs are the gradients; the aggregator is the optimizer.

Prompts, skills, programs, agent harnesses, or your own artifact type. AgentDescent gives each a measurable evolution loop, with interchangeable modules for how candidates are proposed, selected, merged and accepted.

$curl -fsSL https://raw.githubusercontent.com/ThinkFlowLab/agentdescent/main/scripts/install.sh | bash
+
OPEN SOURCE · v0.11.0 · BUILT FOR CODING AGENTS

Evolve
anything.

Diffs are the gradients; the aggregator is the optimizer.

Prompts, skills, programs, agent harnesses, or your own artifact type. AgentDescent gives each a measurable evolution loop, with interchangeable modules for how candidates are proposed, selected, merged and accepted.

$curl -fsSL https://raw.githubusercontent.com/ThinkFlowLab/agentdescent/main/scripts/install.sh | bash
@@ -46,7 +46,7 @@ description: AgentDescent is a modular evolve engine for prompts, skills, progra

Start with your agent.

One command installs AgentDescent and connects supported agents on your PATH.

Open Quickstart ↗
- +