swarm-orchestrator 13.1.9 → 14.0.1

package/CHANGELOG.md

@@ -1,5 +1,322 @@
 # Changelog
 
+## Unreleased
+
+## 14.0.1 - 2026-09-07
+
+### Fixed
+
+- **`swarm ci` no longer charges a patch for a failure the base already had.** Dependencies are
+  installed with `--ignore-scripts`, because install scripts run whatever the registry serves, so
+  a project that builds on `prepare` was verified without its build output and the tests that
+  import it failed. koa routes `import 'koa'` to `./dist/koa.mjs`: 437 passed and 2 failed without
+  the build, 439 and 0 with it, and the two were reported as a regression the patch caused. A
+  failing check is now re-run with the patch reverted on the same checkout, and one that fails
+  both ways is recorded as inherited rather than as a regression, with the build-on-install cause
+  named in the advice. A clean run pays nothing for this, since the base is measured only to
+  explain a failure.
+
+### Added
+
+- **Tasks with two oracles, mined from merged pull requests.** `scripts/mine-pr-tasks.mjs`,
+  `scripts/check-pr-task-viability.mjs` and `scripts/pr-task-pass.mjs`: a merged pull request
+  supplies a base commit, a task, and a specification its own maintainers wrote, whose cases are
+  dealt alternately into a sealed oracle and a held-back one. A candidate is kept only where its
+  tests provably fail on the base source and pass on the merged tree, so nothing is labelled by
+  hand. False-green rate now 0 of 15 across two corpora, 95% CI [0.0, 20.4].
+
+## 14.0.0 - 2026-09-06
+
+### Breaking
+
+- **`swarm ci` no longer reports a patch verified because the project's suite passed.** A suite
+  says nothing broke; it does not say the task was done, because it tests the behaviour the
+  project already had and a task adds behaviour it did not. `verified` now requires both a
+  passing suite and a task oracle that accepts, so a run without `--oracle <command>` exits
+  non-zero with `task: unjudged` and the reason printed. Anything scripted against `swarm ci`
+  exiting 0 needs an oracle passed to it. This is the fix for a measured 22.2% false-green rate,
+  not a tightening for its own sake.
+- **The major version moves because of that one change.** Everything else below is additive.
+
+### Added
+
+- **`swarm ci` reports whether anything broke and whether the task was done, separately.**
+  `regression` says the project's own suite still passes; `task` says a trusted oracle agrees the
+  work was done, and reads `unjudged` where none was given. `verified` requires both, so a run
+  with no oracle is unverified with the reason printed rather than verified by omission. This
+  closed a real 22.2% false-green rate: a repository's suite tests the behaviour the project
+  already had, and a task adds behaviour it did not, so a patch that adds a feature badly still
+  passes it.
+- **`--oracle <command>` supplies that task check.** Nothing infers one. It runs in the fresh
+  checkout after the suite, whether or not the suite passed, so its verdict is readable either
+  way.
+- **`--install` installs the fresh checkout's dependencies from its lockfile**, with install
+  scripts off. Off by default, because installing runs whatever the registry serves. Without it a
+  real project has no test runner in a fresh clone, every check stands down, and the honest
+  answer is that nothing was measured rather than that nothing passed. Reporting the second was a
+  defect, and it is why this exists.
+- **`--isolation <runtime[:image]>` runs commands behind a kernel-enforced boundary**, on docker,
+  podman or nerdctl, with no network, a read-only root, a tmpfs, dropped capabilities, no new
+  privileges, and memory and pid ceilings.
+- **A run measures what it actually executes under, before the first tool call.** The containment
+  self-test runs the escapes rather than reasoning about them: read a host file outside the
+  workspace, write outside it, open a socket. Whatever gets through is named in an
+  `execution-envelope` record and printed. With no kernel-enforced backend the answer is
+  `restricted`, and that is the word used. A backend that refused every escape and could not
+  reach its own workspace is `unknown` rather than `isolated`, since a command that sees nothing
+  is contained and cannot work.
+- **A bundle carries a DSSE envelope over an in-toto statement**, signed over the
+  pre-authentication encoding rather than the payload bytes, so what a run produced can be checked
+  by tooling that never heard of this project.
+- **Durable run state, and the commands to work with it.** `swarm list-runs`, `swarm inspect
+  <run-id>`, `swarm resume`, `swarm retry-step`, `swarm abort` and `swarm repair`, over a
+  `node:sqlite` store in WAL mode with intent recorded before effect and an idempotency key per
+  step, so an interrupted run is taken up rather than started again and a dead run's leases are
+  released rather than held forever.
+- **`swarm gc [--older-than 30d] [--remove]`** says what stored evidence would be removed before
+  removing any of it.
+- **`--json` on a run and on `swarm ci` and `swarm inspect`**: line-delimited JSON, one line per
+  event and one result at the end, each naming its schema.
+- **`confirm_timeout_minutes` under `[interface]`**, so a confirmation nobody answers is refused
+  after thirty minutes rather than holding a run open indefinitely.
+- **The false-green rate is measured against an oracle the tool is never given.** Each
+  real-repository task carries two: one sealed before the runs and handed to the tool, one written
+  from the task text and held back from it. 0 of 11, 95% CI [0.0, 25.9], counting the eleven
+  patches the tool certified. `node scripts/second-oracle-pass.mjs` reproduces it from recorded
+  diffs with no model calls. A single-oracle version of this number was published and withdrawn:
+  the same test on both sides of a comparison agrees with itself.
+- **`scripts/restore-bundle-blobs.mjs` puts offloaded artifacts back**, writing a file only where
+  its content hashes to the digest the bundle names.
+
+- **A run has a wall budget of its own.** `--max-wall-minutes <n>`, or `max_wall_minutes`
+  under `[budgets]` in swarm.toml, bounds the whole run: the first loop and every auto-resolve
+  attempt draw from it, the attempts stop when it is spent, and the run goes on to its final
+  gates and its bundle with the budget recorded on the chain as a `session-budget` record.
+  Before this each loop had its own half hour and the run as a whole had no bound, which is how
+  a run measured inside a forty-five minute container could outlast it and leave nothing: the
+  campaign's first arm lost seven runs that way. Unset, nothing changes: each loop keeps its
+  half hour.
+- **The criteria a run is measured by are sealed on the chain before the model is asked for
+  anything.** Every gate with its severity and the rule that reads it, the budgets, the attempt
+  cap and the ratchet arms, as one record written before the loop, and the embedded verifier
+  holds every gate run to it: a gate the seal does not name, a severity or rule it did not
+  declare, or a sealed gate missing from the final attempt fails verification. A bundle of the
+  new format that carries gate runs and no seal is refused, since that is the one way to dodge
+  the check; a bundle that ran no gates has nothing to be held to.
+- **Every passing gate is bonded.** After the fixed point, each gate that passed is handed one
+  added file it has to refuse, a failing test, a TODO, a credential-shaped token, a file outside
+  the declared set, a change over the budget, and the result is a `gate-bond` record carrying
+  the observation. A check that refused it held. A check that passed over a bond it demonstrably
+  saw is vacuous, which makes a blocking gate's run not green whatever the cycle said, and the
+  exit code says so. A command that passed where nothing shows it read the bond is unshown,
+  named as such and never promoted to held. A gate with no bond is recorded as not bonded. The
+  review page carries the bond beside every gate and paints a vacuous pass red.
+- **A bundle carries `rederive.mjs` beside `verify.mjs`.** Where the verifier asks whether a
+  bundle is what it says it is, the re-deriver asks whether a third party applying this tool's
+  rules would reach the verdicts it records: every gate status from the recorded exit code and
+  output under the rule the record now names, every ratchet decision from its recorded
+  measures, every bond from its observation, every claim from its predicate, and the gate runs
+  from the seal. What it cannot re-derive from the bundle alone, it names rather than agrees
+  with. Dependency-free, like the verifier, and parity-tested against the parsers it mirrors.
+- **Routing by a competency table, class by class, and never by interpolation.** `swarm
+  calibrate` now writes, beside its pick, what each model was measured to do on each class of
+  task: executed repeats and the ones whose gate passed, per sweep, read off the sweep's own
+  `calibration-run` records and folded across sweeps of the same golden set. A task's router
+  asks the table for its class before the reward log has enough samples to say anything, and
+  where a candidate has at least six executed runs on that class the best gate share stands
+  as the default, recorded as a `competency` assignment with the counts it was chosen on. Where
+  no candidate clears the floor, or none has an entry for the class at all, the table abstains
+  by name and the calibration pick stands as before; a model's showing on one class is never
+  read as evidence about another.
+- **A fifty-repository campaign harness, with its criteria sealed before anything was
+  selected.** `campaign/criteria.md` was committed before the first GitHub query, the method in
+  `campaign/methodology.md` before the first run, and `campaign/harness/` sequences the rest:
+  the search saved raw, the candidate walk judged by rules that name the first one failed, one
+  seeded defect per repository accepted only where the repository's own suite passes before and
+  fails after, and one run per arm inside a pinned container on an internal network whose only
+  other member is a forwarder to the model backend. Results are read off the bundle and the
+  tree, never off the run's exit code.
+- **The campaign was run: two local arms over fifty repositories, one hundred runs, ninety-three
+  bundles that verify.** Fifty seeds in five languages, each a repository whose own suite passed
+  before the seeded line and failed after it, run once per arm under a forty-five minute
+  budget. On qwen3.8:27b over rapid-mlx, forty-three runs executed and every bundle verifies,
+  twenty-six of them fixing the seed; seven ran out of budget without a bundle. On
+  qwen3.6:35b-a3b over Ollama, all fifty executed and verify, thirty-nine fixing the seed, at a
+  median of nine minutes against sixteen. The frontier arm was not run, because the key it
+  needs has no balance, and the report carries it with zero runs rather than leaving it out.
+  `campaign/results/report.md` is generated from the result records alone, the bundles are
+  under `campaign/corpus/`, and the method with its dated amendments is in
+  `campaign/methodology.md`. The campaign measured the CLI packed on 2026-09-02 and stayed on
+  it while two defects it exposed were fixed in the tree: a shell allowlist that refused every
+  toolchain but node's, and a model call with no deadline.
+- **Two scheduled proofs on GitHub Actions.** A nightly one runs the suite and the fuzz smoke
+  from a clean clone, then verifies the committed reference bundle with its own verifier and
+  refuses a one-byte copy of it, uploading the transcript. A weekly one installs a pinned
+  Ollama on the runner, seeds a workspace whose test fails on a real defect, runs one task
+  through the built CLI, and verifies the fresh bundle in the same job; the task outcome is
+  recorded and never judged. Neither continues on error, and a scheduled failure of either
+  opens an issue with the run link.
+
+- **A debug copy of what a local backend was actually sent and actually said.**
+  `SWARM_TRANSPORT_TRACE=<path>` writes the raw request body and every raw response frame of
+  every local call to a JSONL artifact, before anything parses them. Off with no path, which is
+  the default: the artifact holds whole prompts and whole completions unscrubbed, so it is not
+  the ledger and not a thing to leave running. It exists for one question the assembled response
+  cannot answer, which is whether a turn that arrived empty was empty on the wire, dropped
+  during stream assembly, or genuinely empty from the model. The response body is teed rather
+  than buffered, so streaming timing is unchanged and the first-token measurement stays real.
+- **A screen for `swarm calibrate`.** A three-model sweep is 180 runs over roughly three hours,
+  the longest thing this tool does, and it wrote plain progress lines and nothing else. It has a
+  view of its own now, because a sweep has none of the things the run screen is built around: it
+  has a grid. Runs finished out of runs planned, the run in flight, a row per model in plan
+  order with green counted over executed runs rather than attempted ones, abstentions named by
+  the reason code the harness recorded, and the last few outcomes. No estimate of time
+  remaining, which would be a prediction people plan around, and no keys, because a sweep has
+  nothing to scroll. Off a terminal it writes one line per finished run in the same words, so
+  the log and the screen stay one account of the run.
+- **An index over the evidence column of the review page.** Records are grouped by turn and
+  folded by default, the per-record digest moved into the expanded payload where the reviewer
+  resolving a claim through it is already looking, and the column opens with how many records
+  there are, how many of each type, a button per type and a search box. The 08-23 calibration
+  bundle is 3,716 records; before this there was no way to find anything in it. A rendering
+  change and nothing else: no record changed, and four existing bundles verify unaltered.
+- **A behaviour probe over changed functions.** A function that answered several ways at the
+  base commit and answers one way now is reported by a blocking gate. Reading the text will
+  never tell a stub from a correct constant, and running it does not have to: the comparison is
+  between two measured variances rather than a judgement about what a function is for. Quiet on
+  a function that was always constant, one that takes no arguments, and one that now refuses
+  every input, which is a tighter signature rather than a missing body.
+
+- **Workers that read what the others have already tried.** They ran side by side and saw
+  nothing of each other, so two could declare the same file and one could spend its whole
+  attempt cap on a gate that had already refused the same fix on the chain next door. They
+  coordinate now through the thing they were already writing: a `read_trail` tool projects
+  typed signals off the peers' ledgers, which files are claimed, which gates failed and how
+  often, which attempts the ratchet rejected, which approaches are spent. No daemon, no bus,
+  nothing anyone writes through, and no read of a chain but the peers' own. Nothing it returns
+  can render green: every signal comes from a record rather than from model text, every line
+  names the peer it is about, and no signal kind reports a success. A run that is not parallel
+  is offered none of it and sends the same tools and the same prompt, byte for byte.
+- **`--redundancy <n>`: try each task several ways and land the best of them.** Attempts
+  diverge because each carries a seed derived from the task, the model and the attempt number,
+  so a report can re-derive them. The winner is chosen by a comparator over harness-measured
+  numbers with the precedence written down, never by asking a model which answer it likes: four
+  earn-it dimensions above every do-less one, a dimension nothing measured abstained on by
+  name, and a dimension only one side measured won by the side that measured it. Counting
+  happens over a file universe fixed across the attempts first, so opening a large test file
+  gains the opener nothing. The whole ranking goes on the chain, losers and their reasons
+  included. A refused winner falls to the next attempt only where the integrated gates were
+  what refused it.
+- **`--goal <text>`: the tool breaks the goal into tasks itself.** A planner run reads the
+  workspace with read-only tools and declares a task graph, which is checked before anything
+  runs for unique ids, resolving dependencies, no cycle, and files that two unordered nodes
+  do not share. The graph is a ledger record written before the first worker starts. Nodes
+  land layer by layer, each layer branching from the tree the one before it left, and a node
+  whose parent did not land is recorded as blocked rather than run against a tree that lacks
+  it. `--tasks` also reads a JSON graph now, so a hand-written one and a declared one are the
+  same artifact. The claim says every declared node ran and landed and says nothing about
+  whether they satisfy the goal, which this tool does not check and says so.
+- **`--concurrency <n>`,** capping how many workers hold a worktree at once. It defaults to
+  one against a local model server, because every worker in the run is asking one resident
+  model for tokens, so the parallelism buys nothing and costs the memory the model lives in.
+  Against a model served elsewhere it defaults to four and deliberately does not grow with the
+  hardware: what the machine is being asked for is N copies of the project's test suite, and a
+  bigger machine does not make those cheaper. The fan-out was unbounded before this, at any
+  redundancy.
+
+### Changed
+
+- **Bundle format 2.** A format 2 bundle carries the sealed criteria, a bond per passing gate,
+  and the re-derivation script; every gate-run record names the rule that read it. Format 1
+  bundles are still read, replayed and verified, and the four committed ones verify unchanged
+  under the new verifier, which reports their lack of a seal as predating sealing rather than
+  as a defect.
+- **The corpus replay reads the v12 falsification corpus from the `v12-final` tag.** It named
+  `main`, which held the v12 tree until `main` was moved onto this lineage; from then on the
+  three replay tests skipped under a green run, on CI and on any clone that had moved its
+  `main`. A tag is the one name that does not move, and the checkout test now pins that the
+  corpus was reached rather than allowing a skip.
+- **The reports the ratchet reads come off streams the harness owns, not files.** The first fix
+  for the counter forgery below had the runner write reports to paths the harness named under
+  the session store. An adversarial pass took that apart: a destination is an argument of the
+  spawned process, so `ps -p $PPID -o command=` hands it to any test that asks, and the file is
+  writable by anything running as the same user, since the writer is the same uid. Getting the
+  harness to read a forgery did not work, because it reads as soon as the runner exits and won
+  that race every time; what did work, four times out of four, was destroying both measures, and
+  an abstention is not a violation. So there is no file: tap goes to stdout and lcov to stderr,
+  and under process isolation a test's own output is captured by the parent and folded into the
+  reporters' streams as escaped comments. The base-control result had the identical hole and
+  gets the identical fix.
+- **The base a run is measured against is resolved to a commit before anything reads it.** HEAD
+  is a symbolic ref spent at the moment each base-side question is asked, and `git` is on the
+  default shell allowlist, so one unconfirmed tool call moved it. Measured on a scratch
+  repository with two of three tests deleted in the tree: after a `git commit -am` the base
+  declared 1 test instead of 3 and the change set held 0 files instead of 1, so the deletion the
+  ratchet exists to catch stopped being a deletion. The parallel path already resolved its base
+  before handing it to any worker; this is the single-run path held to the same thing.
+- **The ratchet's collected and skipped test counts come from a report the runner wrote, not
+  from what it printed.** Node's default reporter passes a test's own `console.log("# tests
+  999")` through ahead of its own counters and the counter reader took the first match, so four
+  print statements in one test file reported 999 collected for a suite of one. That number fed a
+  blocking arm. Where the harness can vouch for the invocation it now asks node for TAP at a
+  path of its own and counts result points there, which a test cannot print into existence.
+  Where it cannot vouch for the invocation both measures are null and the ratchet abstains on
+  them by name, which is stricter than the reading it replaces.
+- **An empty assistant turn is recorded as an abstention rather than as a run of the model.**
+  The harness classifies every turn as it crosses into the ledger and stamps the verdict on the
+  record, so a repeat whose only turn arrived empty reports `abstained` with a machine-readable
+  reason instead of being scored against the model as a wrong answer. Two calibration bundles
+  had been scored with those folded in. Calibration now reads `executed` off the records rather
+  than off the loop's own counter, because a reviewer re-deriving that number has the records
+  and not the loop.
+- **A seed is recorded with whether the backend took it.** Ollama and rapid-mlx both accept a
+  seed field and neither promises to sample from it, and the SDK reports a refused setting as a
+  warning. Those warnings now reach the record, so a seed in a bundle is never read as a seed
+  that made the run re-derivable.
+- **The derivation heuristic reads a shell command as a command.** Its threshold is unchanged,
+  because lowering it flags ordinary commands that mention a filename someone read. What changed
+  is that where an argument parses as a shell command it is also compared as one, with flags
+  dropped and interpreters folded together, so inserting `-fsSL` and swapping `sh` for `bash`
+  no longer rewrites it past the match.
+- **An assertion comparing a value with itself stops counting as an assertion.** The rule was a
+  literal against the identical literal; it now substitutes the file's own bindings first, so
+  `expect(v0.a).toBe(v0.a)` and the spelling that binds one side to a name are both seen. This
+  is arithmetic over expressions and not a judgement about meaning: deciding that two
+  *different* expressions mean the same thing is still refused.
+- **The secret-scan gate reads a credential written in pieces.** Where a change rejoins them,
+  by concatenation or by a template literal, the gate rebuilds the value those lines make and
+  hands it to the detector that already decides about credentials, under the name the change
+  gave it. No second detector and no new threshold. Halves that are never rejoined build no
+  value and are still not caught, which build guide 7.1 says in those words.
+
+### Fixed
+
+- **Every evidence bundle verifies from a clone again.** An earlier weight reduction moved record
+  payloads out of the tracked tree, which stopped 47 of 51 bundles passing their own verifier,
+  four of them cited in the README and `claims.md` in words like "verify.mjs exit 0". Payloads are
+  never offloaded now; only derived artifacts are, and those are regenerated from the records.
+- **The README no longer cites a number the documents behind it withdrew**, and its gate tally is
+  counted from the table rather than from memory.
+
+- **A parallel run now sweeps up the branches it created.** Worker branches outlive their
+  worktrees on purpose, because the merge queue merges from them after the working copy is
+  gone, and nothing outlived the queue: a repository gained a branch per worker per run, for
+  ever. The integration branch is never swept, since that is the result. It prunes worktrees
+  first, which is what stops a run killed part-way from making the next one fail on a path git
+  still believes in.
+- **`--goal` says how the planner stopped, not just that it declared nothing.** Running out of
+  steps, answering in prose, and returning nothing at all want three different responses, and
+  the message treated them as one. It now names the stop reason and the step count and says
+  what to try for each.
+- **Two review-page summaries that trailed off after the colon.** `worker-finished` and
+  `merge-attempt` both asked for a field neither writer emits, so they rendered as
+  `worker-1 finished:` with nothing after it. They read what is actually recorded now.
+- **The weekly scan stopped filing the same 21 findings every Monday.** Data is excluded by
+  path, meaning the secret scrubber's own fuzz corpus and the captured shakedown logs and
+  nothing else; the nine findings in real source carry a suppression naming the one rule and the
+  reason, at the line that carries the finding. Under the workflow's own command it now reports
+  zero and exits 0, so the weekly issue means something arrived that nobody has looked at.
+
 ## 13.1.9
 
 ### Added