Commit Graph
108 Commits
Author SHA1 Message Date
jason.woltje 6ad11e6993 fix(pi): pi-unpin S3 re-check 10: unshare's outer layer is best effort: each fit check probes it once, alone and then with bwrap inside, and runs without it when unshare is missing, older than util-linux 2.36 or refused (Ubuntu 24.04 with apparmor_restrict_unprivileged_userns=1), with bwrap's argv the same either way and no refusal; only a failed bwrap or a missing setpriv refuses, each named; the sweep stops processes of this uid whose command line names a stale scratch dir before it removes it, scratch names carry the pid namespace and a sweep skips other namespaces' dirs, and a run or probe the check kills is swept at once; test 655 waits for pi.ready() before its fixture starts the launcher (#1873) 2026-11-20 01:17:33 -06:00
jason.woltje e52a6b714f fix(pi): pi-unpin S3 re-check 9: a fit check dies with the process that runs it: its command line runs under setpriv's parent-death signal and a shell that checks its parent is still that process, then unshare puts bwrap as pid 1 of a fresh pid namespace with --kill-child, so a process bwrap leaves in its setup dies with the namespace, and bwrap gets --die-with-parent; each run's scratch directory carries its process's pid and start time and a check first removes those a process no longer running left; the fit route is zai glm-5.3-flash, with a test that goes red on a Claude model or Anthropic provider; the launcher's warm-up records a harness plugin it can't load once, by code only; test 655 waits for pi's ready() before its timed part, with its own timeout raised (#1873) 2026-11-20 01:17:33 -06:00
jason.woltje 2d065b936e fix(pi): pi-unpin S3 re-check 8: the launcher starts pi's fit check in the background when it starts, through one harness load helper a launch shares, so the first launch joins the check in flight; nothing waits on it, a failed warm-up is dropped with no cached verdict and no unhandled rejection; a fit check that throws refuses as harness-feature-missing with no path in its line; tests for the launcher's rejected preflight and the Triage host's wait on ready(); the fit's provider and model move to the pi module's session-env.mjs (#1873) 2026-11-20 01:17:33 -06:00
jason.woltje 098b3771c0 fix(pi): pi-unpin S3 re-check 7: the fit check runs its pi processes with async spawn, four at a time, and every caller awaits it, so the bus's event loop stays live (longest gap 22 ms against 6031 ms before) and two launches during one check share it; the launcher's call() awaits a module's promise and refuses its rejection as a throw, and preflight and version are awaited; the Triage host and route await the module's ready(); a run whose pipes outlive it settles a second after it exits or is killed; the fit's host env is the process's own through sessionEnv minus provider keys, its names are in the cache key, parity runs over a login shell's host and a unit's, and stand-ins that leak with TERM, TZ or LC_ALL unset are refused; bundle tests await harnessReady() (#1873) 2026-11-20 01:17:33 -06:00
jason.woltje eac42663ba fix(pi): pi-unpin S3 re-check 6: the fit runs get a bus session's environment, built by the launcher's sessionEnv over ENV_ALLOW, the sandbox's sessionSetenv and the runner's adapterEnv with the pi module's harnessEnv (now in session-env.mjs), so no run sets PI_CODING_AGENT_DIR and the agent sources sit at $HOME/.pi/agent; a test fails when the fit env and a bus session's env differ by a name outside one listed exception; the drift test runs the adapter in that env over 98 argv sets, each tool count, -c case, thinking level and provider, and checks every variable the adapter reads; -c over a session dir with nothing for the project is named under Not run, with a gap test (#1873) 2026-11-20 01:17:33 -06:00
jason.woltje b41bc00b26 fix(pi): pi-unpin S3 re-check 5: five -p runs in a bus session's shape pass every pair of the adapter's session branches (tool list or --no-tools, -c over a planted session, --thinking, --mode json), a -c run must resume the planted session, each loads two -e, the -e extension's command list shows skills and prompt templates in -p, the drift test drives the adapter through all sixteen branch sets and fails on a pair no run passes, and a test holds the --no-mcp run on its own (#1873) 2026-11-20 01:17:33 -06:00
jason.woltje 749d6d98f5 fix(pi): pi-unpin S3 re-check 4: one sealed run in a bus session's shape, built by sessionShape, which a test holds to the adapter's argv; a one-out run must show only its own kind; the README and header list what a session passes that no fit run does (#1873) 2026-11-20 01:17:33 -06:00
jason.woltje fa5ce3ae65 fix(pi): pi-unpin S3 re-check 3: the control passes --no-tools and --no-themes, one sealed run per loader flag leaves only that flag out and its kind must show, every run gets a fresh scratch directory, and the README says what argv the check runs (#1873) 2026-11-20 01:17:33 -06:00
jason.woltje d009ccaec4 fix(pi): pi-unpin S3 re-check 2: the control passes the sealed runs' -e and --system-prompt so the runs differ only by the seal flags, a project source no longer shows a kind, a run over 4 MB says so, and --version gets 3 s (#1873) 2026-11-20 01:17:33 -06:00
jason.woltje 2613fb1c4c fix(pi): pi-unpin S3 re-check: the control counts a kind only through a source the sealed runs can show, so a SYSTEM.md no longer stands in for the project APPEND_SYSTEM.md or an AGENTS.md (#1873) 2026-11-20 01:17:33 -06:00
jason.woltje df26085303 fix(pi): pi-unpin S3 review: the control must show each planted kind and answer get_commands, the -e extension reads the system prompt so context and trusted project files are checked, settings and package sources are planted, help flags come from the option column, missing -e and the seal run in -p too, and the digest covers npm's lock files (#1873) 2026-11-20 01:17:33 -06:00
jason.woltje 8908df113e feat(pi): pi-unpin S3: every start runs a sandboxed, offline fit check on the pi it found and refuses harness-feature-missing when a flag is missing, a missing -e starts, or a planted source loads with the seal flags on (#1873) 2026-11-20 01:17:33 -06:00
jason.woltje 07eca4a670 fix(chat-fail): review round 3: the link, email and mixed-run redaction patterns start only where a run without spaces starts, so redacting the whole kept stdout stays linear (a 128000-character run took 10 to 15 s, now a few ms) and matches the same text; a long text that redacts below the read limit is not cut and keeps its first word, and a cause that is only [redacted] is dropped; the proxy page reads relogin, plural and past forms of an expired sign-in as a sign-in (#1873, lane chat-fail) 2026-11-19 16:36:16 -06:00
jason.woltje fc9e4dfca8 fix(chat-fail): review round 2: the F1 cause is redacted over the whole text before the tail cut, so a cut that lands between a key word and its value no longer shows the value; the account sign-in words in the proxy page match whole words only ("assigning" and "logging" no longer read as a sign-in) and a prefixed or "is expired" token counts (#1873, lane chat-fail) 2026-11-19 16:36:16 -06:00
jason.woltje 10a35ce65a fix(chat-fail): review round 1: the F1 cause decodes %XX, HTML references and NFKC forms before the secret and redaction checks, redacts the value after a quoted or space-separated key word and Windows and UNC paths, and is read from the tail of the output, with the cause on the retry log line too; a launch-failed start posts only when the same message is refused twice, project-pending and processing-halted are waits, and a refused post is tried again and logged once; the proxy reason refuses an impossible date, checks quota_blocked before the message words and no longer reads a bare "expired" as a sign-in (#1873, lane chat-fail) 2026-11-19 16:36:16 -06:00
jason.woltje 72ea7272ec fix(harness): chat-fail F1: the turn-failure code moves into the mounted harness failure-notice.mjs, so a sandboxed runner loads it; the separate file was outside RUNNER_FILES and every sandboxed run died with ERR_MODULE_NOT_FOUND (#1873, lane chat-fail) 2026-11-19 16:36:16 -06:00
jason.woltje 27a359590e fix(harness): chat-fail F1: a failed turn's chat line carries the harness's error as a plain cause and next step, cleaned, redacted and capped, with the technical reason under it (#1873, lane chat-fail) 2026-11-19 16:36:16 -06:00
jason.woltje a3ad6578c4 fix(harness): review round: a legacy model entry (a provider or route name, no module) carries no account data, so a Claude-named model on it is treated as OAuth and held to claude-code; the Triage host refuses a key in any bundle/ file and any URL but the proxy's in a harness file; the models.json comment says what route() can and cannot see (#1873, lane harness-matrix) 2026-11-19 14:43:26 -06:00
jason.woltje cd58ac23cd feat(harness): an OAuth Claude model runs on claude-code only; Triage runs on claude-code or pi through the proxy (#1873, lane harness-matrix)
cliproxy holds a model to claude-code when its upstream is claude, when an OAuth Claude account
offers the same name, or when the name contains "claude"; every other proxy model runs on
claude-code and pi. The catalogue, the launch (sessionModel and the models.json list) and the
provider route apply the one rule and refuse with model-harness-incompatible, so a stored
Claude-on-pi route is refused, never rewritten. The other route's pi provider is mosaic-proxy-other.

Triage: the route asks the provider for the transport on the stored harness and holds the
harness's env() to one contract (the key, unchanged, under exactly one of ANTHROPIC_AUTH_TOKEN or
ANTHROPIC_API_KEY; the URL only as ANTHROPIC_BASE_URL), with no list of harness names. The host
gives the module the real proxy and transport, checks each file it writes under the session's HOME
(regular, one link, no key) and binds it read-only after the run directory, and refuses a route env
with no key name or two. PATH starts with node_modules/.bin, as a bus session's does.

Tests: claude-code runs a z.ai and a codex proxy model against a stand-in (real claude binary);
pi answers Triage through a stand-in under bwrap; every new fence has a test.
2026-11-19 14:43:26 -06:00
jason.woltje 387f73a7a4 fix(harness): r127 review round, only a shell's own not-found line names a program (#1873, lane r127)
Review P3-1: a "No such file or directory" line from another program
(cat, ls) no longer counts. Only bash's "line N:", dash's "N:" and env's
quoted form do.
P3-2: a name with a run of 16 letters and digits looks like a key, so it
stays in the run's private stderr instead of the reply.
P3-3: a 16 KiB stderr window that starts mid-file drops its partial first
line, so a cut name ("aude") isn't reported.
P3-4: CRLF lines read the same as LF lines.
P3-5: tests for the name check (tokens without spaces, key-shaped names),
for a cat line, and for the window boundary through the runner.
2026-11-14 21:19:02 -06:00
jason.woltje 61371adb10 fix(harness): an adapter exit 127 names the program that wasn't found (#1873, lane r127)
A claude-code turn on a Console started without MOSAIC_CLAUDE_BIN failed
with a bare "adapter exited 127". The sandbox binds /usr, so the wrapper
/usr/bin/claude is visible, but its exec target /opt/claude-code/bin/claude
isn't bound, and the shell exits 127.

On exit 127 the runner now reads the tail of the turn's stderr, finds the
shell's "not found" line (bash, dash, env) and adds the program's basename:
"adapter exited 127 (command not found: claude)". It never shows a host
path or anything that isn't a plain program name; a 127 without such a
line keeps the bare reason.

The deploy side (MOSAIC_CLAUDE_BIN on the Console unit) is a Lead-owned
script change, written up in the lane report.
2026-11-14 21:07:14 -06:00
jason.woltje 1ced0e3924 fix(harness): escape every control and format character in workflow data (starter packs SP7 review P3)
The workflows data lines now escape every \p{Cc}, \p{Cf}, \p{Zl} and
\p{Zp} character by UTF-16 unit, so zero-width, soft hyphen, BOM and
tag characters never reach the prompt invisible. The section says the
rails come later in the prompt, not at its end, since skills can follow
them. A role holding task.assign or task.create only cross-role sees
every step; now tested.
2026-11-12 16:00:42 -06:00
jason.woltje 4bd260457e feat(harness): workflow guidance in the prompt (starter packs SP7)
resolveInstance carries the organization's workflows, plus the given
project's, the project's winning by name; the key is added only when
there is one, so an organization without workflows digests as before.

The prompt renders them under one fixed heading, "## Organization
workflows", after the contract and persona and before "## This
session". The chief of staff and roles holding task.create or
task.assign see every step; a role a step names sees only its own
steps; any other role sees nothing. Each workflow is one JSON line
between <workflows-data> markers with <, >, &, line separators, C1
controls and bidi overrides escaped, so no title, description or step
role can start a heading or close the block. The section says
workflows grant nothing and the contract and rails win. The contract
text is unchanged.
2026-11-12 15:47:32 -06:00
jason.woltje 87db40e0cd fix(cli,seat,discord,organization,control-board,harness,plugins,webui): review round: launcher reasons, session, discord reads, mosaic-config and keys by label (#1873)
Review "changes 8" fix round for org-redact6.

- launcher: a harness module's throw names its system code or class,
  never its message; preflight, loader, bundle, checkout, sandbox and
  route text is relayed with host paths replaced (relay()).
- claude-code and pi preflights name MOSAIC_CLAUDE_BIN and the
  repository's node_modules, never a path; persona and skill bundle
  errors and sandboxProblem name their file or the sandbox by label.
- seat session: thread store, legacy store, overlap, busy lock and
  newRun refusals name the place under the data root.
- discord: unknown keys, binding names and turn ids are described when
  not plain; approvals, journal and repository read errors name the file
  and code; git drops stderr lines that carry a host path and a spawn
  failure names its code only.
- seat, control-board: registration fields and seen keys are described
  when not plain; invalid seat name and layout state the constraint.
- organization: the registry names the config directory and registry
  file by label; --project is quoted only when it is a plain name;
  systemConfigProblem moves to its own module.
- scripts/mosaic-config.mjs: the config by where it came from, the data
  root and its marker by role, keys and arguments only when plain, and
  error codes in place of raw messages. Exit codes unchanged.
- webui sso-config: an unknown console key is named only when plain.

Sentinel tests in each package's path-redaction-labels file fail on
BASE; pins that quoted a path now assert the label and its absence.
2026-11-11 08:53:31 -06:00
jason.woltje 5e93356cfa fix(authority): close chief grant and root fire review gaps (#1873) 2026-11-11 04:19:33 -06:00
jason.woltje 89205a6163 feat(authority): resolve hire, fire and role changes through pinned human decisions (#1873) 2026-11-11 04:19:33 -06:00
jason.woltje b44b14c7df feat(plugins): install pinned instance packages and declared tools (#1873) 2026-11-11 00:11:13 -06:00
jason.woltje c8ff8ea128 fix(redaction): omit remaining configuration values from refusals (#1873) 2026-11-10 19:56:05 -06:00
jason.woltje 7035ce61b3 feat(harness): expose caller rails and refusal guidance (#1873) 2026-11-10 18:51:15 -06:00
jason.woltje 307fe928ac fix: redact remaining file and config refusal echoes (#1873) 2026-11-10 17:06:54 -06:00
jason.woltje 6f0c72a02f fix(organization): redact model selections and JSON parser diagnostics (#1873) 2026-11-10 15:39:53 -06:00
jason.woltje a8d9eecf65 organization: move packages/business to packages/organization (rename r2 stage B1)
git mv packages/business -> packages/organization (@mosaic/organization). The
old package stays as one-line re-export shims of each moved module, with
src/business.mjs re-exporting src/organization-config.mjs and src/cli.mjs
forwarding to the moved executable. BusinessError is OrganizationConfigError
(one constructor, name kept BusinessError); OrganizationError stays distinct.
The bus adapter gains organization-context.mjs (busOrganization), with
business.mjs (busBusiness) as the same function.

Import consumers, manifests, the lockfile, scripts/mosaic, stack-copy.sh
COPY_PATHS (both roots), seat sandbox RUNNER_FILES and the copy closure
walker (organization-imports.mjs, both roots required) move to the new path.
v1 serialization, v1 writers and the stage A fixtures are unchanged; the stage
A contract tests stay where they were and pass as before.
2026-11-09 11:47:48 -06:00
jason.woltje 623388603b Merge organizations MVP (ui-org-r2) into refactor
BUILD-LOG.md: append-only union, refactor entries first, then the branch's.
docs/SESSIONS.md: append-only union, refactor entries first, then the branch's.
packages/bus/src/runtime.mjs: keep the organization usage timer and refactor's halted/invalid startup-recovery gate.
packages/harness/src/first-output.mjs: keep the branch's harness-neutral markers; refactor's empty-length rejection moves into the Pi monitor (packages/plugins/harness/pi/model-stream.mjs).
packages/harness/tests/runner.test.mjs: keep refactor's length-continuation cases and the branch's output-policy cases.
packages/plugins/harness/pi/index.mjs: keep the documented GLM limits and the branch's transport-based bundle.

Suites: 6356 pass, 32 fail; all 32 fail identically on both parents
(node_modules symlink EISDIR in pinned-copy and stack copy/release tests,
missing Python jsonschema in two conversation tests).
2026-11-08 21:38:42 -06:00
jason.woltje b48396b68e fix(plugins): provider review fixes, admission binding, alias-safe models, no provider names in core (#1858) 2026-11-06 19:52:57 -06:00
jason.woltje ccccac2ade feat(plugins): model providers as plugin modules, cliproxy catalogue on Routing (#1858) 2026-11-06 19:11:13 -06:00
jason.woltje 08b7db413b fix(harness): move harness-specific output handling into the harness plugins (#1855) 2026-11-06 18:27:58 -06:00
jason.woltje 3e03e123fe fix(harness): check processing and close usage on every length continuation path (#1852)
Review round fixes for the length continuation in Part A.

Before each continuation the runner now reads processing.status as well as
state. A halted, invalid or unreadable processing state starts no further
attempt; the message is left unread with no reply and no failure record, as
for a halt during any turn.

A failed state or processing read keeps the empty length classification, so
the sender gets no RESULT. The chief notice and the trail event name how many
attempts hit the limit.

usage.end for the message id now runs in a finally block, so it closes once
even when an attempt throws.

The library citation retry no longer starts continuations. They would run
outside the message's usage bracket, and the retry's own bracket is refused
(#1853). An empty length retry fails visibly instead, with no fixed RESULT.

Tests cover a halt between attempts, a failed continuation check on the plain
and library paths, a throwing attempt, a library continuation that succeeds,
and the retry path.
2026-11-06 16:34:04 -06:00
jason.woltje 3c689efa31 fix(harness): treat a length stop with no answer as a failed turn (#1852)
Pi could spend its output limit on thinking and the runner relayed an empty answer as a result. Preserve final stop metadata and allow two bounded continuations, accounting for the whole message in one transcript usage delta.

Record exhaustion on the message trail and notify the chief without an automatic reply cycle. Mount the shared notice prefix in the sandbox.
2026-11-06 16:12:30 -06:00
jason.woltje 4048e0f60c fix(bus,harness): stop INFO echo replies and duplicate continuation notices (#1847, #1848)
Expose the parent message class on agent delivery so the runner can stop automatic replies to answers while preserving requests and ordinary relays. Scope continuation retry versions to the target, suppress notices during processing halts and unclaimed launches, and clarify the instance-running refusal.

The bus suite excludes project-files.test.mjs because its 1 MiB socket test timed out on both the candidate and the untouched 1186cf7b1 base.
2026-11-06 14:14:53 -06:00
jason.woltje 97d670da94 fix(bus): name task evidence gaps and stop request polling (#797) 2026-11-06 10:49:33 -06:00
jason.woltje a7417a1f2e fix(project): accept JSON strings for records 2026-11-06 07:40:28 -06:00
jason.woltje 8c568ea467 Merge integration head into #797 stall guard 2026-11-06 03:49:38 -06:00
jason.woltje 6b2c925b95 fix(harness): fail 1 on an unmapped child signal (#797)
A signal name with no number in os.constants.signals now exits 1 instead of a made-up 128. Known signals keep 128+n (SIGKILL 137, SIGTERM 143) and the content-free stderr receipt. Test covers all three through the real entrypoint.
2026-11-06 03:22:31 -06:00
jason.woltje a3f4551406 fix(harness): back off only running instances (#797) 2026-11-06 03:01:12 -06:00
jason.woltje 006381cfc4 fix(harness): retain model child signal receipts (#797) 2026-11-06 03:01:12 -06:00
jason.woltje 478df1cacd fix(harness): preserve adapter text output bytes (#797) 2026-11-06 03:01:12 -06:00
jason.woltje 3e754c96a0 fix(harness): back off refused launches (#797) 2026-11-06 02:05:32 -06:00
jason.woltje c7ddc7d31c fix(harness): mark first streamed model output (#797) 2026-11-06 02:02:43 -06:00
jason.woltje 27315d4961 Merge fix/mvp3-coder-stall-797 (8d27a82ea) into integrate/mvp3-ux-1106
Gate on Node 26.12.1, 31 files: 422 pass, 1 fail. The failure is the
row 66 unslop timing flake in packages/harness/tests/claude-session.test.mjs,
which fails about 1 run in 3 on the base too; the file rerun passed 14/0.
Append-only BUILD-LOG.md and docs/SESSIONS.md conflicts keep both sides.
2026-11-05 22:39:56 -06:00
jason.woltje aca8ce0599 Merge fix/mvp3-planning-797 (c47a67de8) into integrate/mvp3-ux-1106
Gate on the merged tree: business 307/0, bus 1596 pass with one
pre-existing hang (project-files.test.mjs, same on relay base 2f1cef4c1),
harness 243/1 on the row 66 unslop test, which also fails 1 in 3 runs
on the relay base (timing flake, pre-existing).
2026-11-05 21:44:04 -06:00