Mosaic concepts pages now own the adapted content; source/license
metadata under docs/reference/concepts. Adds ACT-1 agent-context
planning capture, pinned concept test package + preparation utility,
foundation observation notes (durability, evidence, federation,
onboarding, workflow), and the #1495 consolidation assessment.
TOOLS.md updated for the host-dev launcher.
scripts/agent.sh --host-dev delegates to scripts/agent-host-dev.sh;
agents/darkwing/launch.sh provides the native development TUI
(context, skills, coding tools, /goal). Root SOUL.md is the M14-era
default-collaborator persona captured by the ACT-1 context work.
Launcher regression checks pass (test-darkwing-launch.mjs).
Rocko-authored, Filbert-reviewed inspector (r6 manifest
a4a44930...) with full review/build/verdict evidence under
docs/plans/reviews. 43/0 selftests, oracle zero-disagreement,
foundation checker PASS. Owner A9 acceptance recorded separately.
Owner adjudication:
- canonical harness IDs match executables: pi, claude, codex, opencode
- agent.json uses one scalar harness ID; registry/manifest resolution, no
hard-coded schema enum
- target mosaic harness list/detect/install/rm/status lifecycle
- detection recognizes reviewed executables and records compatibility
without reading/copying harness homes
- installs are exact-version/verified, Mosaic-managed, never global
- detected external harness is available but not container-ready until
imported/installed, absent a separately reviewed host adapter
Gate 1 resolved. Remaining P0/review gates stay open; no implementation
authorized. Suites 24/15/90/14/17 + verify green; unslop clean.
Owner direction: the stack must never impact default harness usage.
Correction to M19 as shipped (nothing had been created in ~/.pi — the
move breaks nothing).
- Mosaic-managed accounts: <dataRoot>/auth/<account>.json, perms 0600
enforced (loose perms flagged in listings, refused by --auth — mirrors
gitea-api.sh credential hygiene).
- ~/.pi is read-only to the stack, permanently; the only interaction
remains the existing read-only container mount of the default
credential. Recorded as a ROADMAP standing decision.
- auth.sh is now config-driven (data root from config.json, fail closed,
consistent with every other tool); status reports both sources labeled.
- agent.sh --auth resolution moved after load_config (needs the data
root); missing/symlinked/non-0600 accounts refuse.
- test-auth.sh: 15 no-Docker cases (accounts-create-nothing, loose-perms
refusal, invalid-config refusal added). Test-authoring correction
recorded in BUILD-LOG (fixture-state mismatch caught before running).
Suites 24/15/90/14/17 + verify green.
Owner decision after live verification of M18: an explicit agents-dir
override that cannot resolve the named seat now refuses the launch
(exit 4, names the seat and dir) instead of launching seatless and
unbounded. Unsetting the override keeps the M13 plain governed TUI.
MOSAIC_ROLES_DIR needs no symmetric change - the M18 gate already
refuses unresolvable role contracts.
Task suite 88 -> 90 (refusal + refusal-names-the-seat). TOOLS.md Agent
section documents the refusal.
Suites 24/90/14/17 + verify green.
Role contracts (roles/<role>.json): roleVersion, name bound to filename,
tools ceiling (subset of pi built-ins), network declared (none|api-only|
open; enforced when network policy lands). Strict schema, fail closed -
a non-role document refuses resolution.
mosaic-task.mjs resolve-role: config-free contract validation, emits
MOSAIC_ROLE_TOOLS / MOSAIC_ROLE_NETWORK.
agent.sh: a declared role binds to its contract. Missing/invalid contract
refuses the launch (exit 2, names the role - the under-equipped-seat
failure mode, mirroring M17 skills). Effective tools = ceiling ∩ requested
(CLI --tools or agent.json caps); no request -> ceiling stands; narrowing
and tool-free outcomes loud on stderr. Adapters unchanged; headless M9
chain (mission ∩ task) untouched.
Ships roles/researcher.json (existing seat declares the role; without the
contract the fail-closed gate would refuse its launch).
Task suite 74 -> 88: contract resolution, wrong-kind/name/network/
duplicate/unsupported/missing refusals, ceiling narrowing E2E (mock
adapter), tool-free E2E, missing-contract refusal. Test-authoring
correction recorded in BUILD-LOG (a check that registered on one path
only, caught by count arithmetic).
Suites 24/88/14/17 + verify green.
ms-communications: integrated as-authored - owner preamble restructure +
collaborator delivery-discipline hunks from the #43 calibration (own
session output is not a send path; the tool performs the preamble flip;
receiving rule 3 requires actually running agent-send.sh).
ms-conductor: collaborator redraft integrated (canon-aligned tracking
surfaces, one-action cadence, fail-closed core) with one conductor
remediation - step 3 now distinguishes refusal (fail closed, never
bypass) from runner outage (direct dispatch to a qualified live seat via
ms-communications permitted, recorded loudly as degraded: no sandbox, no
run record; suites still gate integration). Preserves the owner's
outage-dispatch intent inside invariant 6.
docs/TOOLS.md: release.sh ensure row added (M16 subcommand existed in
code but not in the doc - flagged by the collaborator, verified in
release.sh usage).
Authorship: owner (ms-conductor doctrine, preamble restructure) +
ms-test collaborator (delivery hunks, redraft); remediation + integration
by conductor (dragon-lin:darkwing). Suites 24/74/14/17 + verify green;
unslop clean.
Collaborator-authored via conductor-loop calibration: task dispatched to the
live ms-test seat (glm-5.3-flash) over agent-send.sh; diff reviewed line by
line and every documented flag/exit code independently verified against tool
source by the conductor; suites green at integration (config 24 / task 74 /
release 14 / conductor 17 + verify).
- new 'Tools (host-side)' section: agent-send.sh, agent-watch.sh, unslop-check.js
- intro reading guide now points at tools/ (worker-flagged addition, accepted)
- Maintenance suite counts corrected: test-task.sh 58 -> 74
Authored-by: ms-test collaborator (glm-5.3-flash)
Integrated-by: conductor (dragon-lin:darkwing)
- scripts/skill.sh: install (bundled or path) / activate / deactivate /
uninstall (refuses while enabled) / list
- skills-enabled + skills-available dirs under the data root; a skill not
in skills-enabled is not enabled or available for use
- pi adapter: MOSAIC_SKILLS -> --skill per dir; --no-skills when none
- agent.sh: seat definitions declare skills[]; resolution against
skills-enabled refuses the launch loudly when missing
- ms-* skills completed (owner-authored canon, hands-off): ms-tools
adapted to the runtime, ms-file-read/write/agent/conductor bodies
written in the owner's style; ms-agent-watch + ms-unslop untouched
- tasks/USER.md onboarding fixtures; suite hardening (nested def path,
user seed, mock-adapter dispatch evidence)
Suites: config 24, task 74, release 14, conductor 17, verify PASS.
RELEASE 0.0.12 packaged; health-gated activation on merge.
Closes#40, closes#41, closes#42
- release.sh health gate runs with MOSAIC_ENSURE_SKIP=1: the gated task run
cannot re-enter release self-determination
- run-task.sh warns on release drift instead of silently using a stale image
- ROADMAP: M20 decision recorded (packages/* monorepo at usurpation,
continuity-first); restructure sequenced as M20 phase 1
Closes#39
- scripts/agent.sh <name>: launches interactive pi TUI in the container
with contracts + optional mission + agent identity + named session +
optional workspace/tools; the Mosaic alternative to vanilla pi
- pi adapter: MOSAIC_INTERACTIVE branch (clean TUI, no -p, no initial
prompt); headless exec rebuilt via positional args (no word-splitting
on the request); MOSAIC_AGENT_NAME optional in headless
- loader: AGENT IDENTITY section when the launcher names the agent
- compose: fixed command removed (request defaults live in run-agent.sh);
MOSAIC_INTERACTIVE/MOSAIC_AGENT_NAME passthrough
- docs/TOOLS.md: full on-demand tool reference; AGENTS.md routes to it
- RELEASE -> 0.0.8 (container change); build verified
Closes#35
Owner direction: the repository root holds first-class, bootstrap-required
configuration only. conductor-policy.json is a ROLE contract (the
conductor's authority), one of scores of future role contracts
(agent-policy, coder-policy, ...) - such files get a dedicated home.
- roles/conductor-policy.json (git mv)
- conductor-apply.sh + test-conductor.sh read the new path
- CONDUCTOR.md records the roles/ convention
Closes UX follow-up from owner layout review; no issue (convention change).
- mosaic-task.mjs prune [--keep=N] [--yes]: default keep 50; without
--yes lists candidates without deleting
- only r-* directories under the runs root; symlinks skipped;
sessions/workspaces/state/config untouched (asserted by suite sentinels)
- append-only receipt runs/.pruned.log records every pruned id
- test-task.sh: +8 retention cases (dry-run no-delete, keep-N, newest
kept, receipt, isolation, invalid keep, empty no-op)
Also: suite hardening - prune section scopes its config per-command
(no export/unset leaking into later sections); duplicated check()
removed; latest_reason hoisted to helpers; status colors now green OK /
red FAIL (terminal-only, NO_COLOR-aware) per owner UX feedback.
Closes#32
Owner feedback: grep match-highlighting made the word 'policy' red while
status words were plain - counter-indicative. Suites + verify now emit
ANSI colors (green success, red failure) when stdout is a terminal;
piped/machine-parsed output stays plain, honoring NO_COLOR. Word 'ok'
promoted to 'OK' for scannability.
Verified byte-level via forced-pty run; piped output unchanged; suites
41/24/14 + verify green.
- mission schema: optional capabilities.tools (same validation as task)
- merge semantics in runTask: neither -> none; mission only -> mission;
task only -> task; both -> intersection (task narrows, never widens);
empty intersection -> tool-free run with an explicit stderr note
- result.json records EFFECTIVE tools; task/mission snapshots remain the
immutable declaration of intent
- adapters unchanged; host-side only (no image change, 0.0.6 still active)
- task suite +5 cases (41 total): all four merge cases asserted from run
evidence + invalid mission capabilities rejected
Policy decision recorded: missions govern; tasks cannot escalate.
Closes#30
result.json recorded lineage correctly; the human-facing show command
omitted the field. Found by owner test: show | grep retriedFrom was
empty on a run whose result.json contained it.
Closes#29
- retryRun rewrites a snapshot's relative mission path to the run's own
recorded mission.json (absolute) before execution — retries stay
faithful to what originally ran
- runTask accepts options.retriedFrom; retry records lineage in
result.json (additive optional field, no schema break)
- task suite +4 cases: retry succeeds, lineage recorded, mission section
present after retry (36 total), missing-run retry exits 4
Closes#28
Collaboration record (conductor loop, docs/plans/CONDUCTOR.md):
- round 1 (worker session worker-1, 2m28s): retry implemented per spec
- conductor live test exposed spec gap: direct invocation lacked
launcher env exports
- round 2 (same worker session, 59s): spawnEnv made self-sufficient,
but used PI_* where compose interpolates MOSAIC_*
- conductor hotfix: 3-line rename to MOSAIC_PROVIDER/MOSAIC_MODEL/
MOSAIC_DATA_ROOT
Final: node scripts/mosaic-task.mjs retry <runId> re-executes a run's
task snapshot as a new run; live retry replied REMEMBERED; all suites
green (24/32/14 + verify).
Known limitation: retrying a run whose task used a RELATIVE mission path
resolves it against the temp dir; lineage tracking deferred.
Closes#25, closes#26, closes#27
- task schema: optional workspace (absent | :run ephemeral | named
persistent under dataRoot/workspaces) and capabilities.tools (pi
documented tool allowlist); strict validation, traversal-proof names
- runner: creates host workspace, passes MOSAIC_WORKSPACE (container
path) + MOSAIC_TOOLS; result.json records both
- pi adapter: cds into workspace; --tools when allowlist present else
--no-tools
- mock adapter: logs delivered MOSAIC_* vars to stderr as deterministic
plumbing evidence (dash prints 'export K=v', so use env not export)
Closes#20
- load-contracts.sh: MOSAIC_MISSION_FILE (readable) appends a MISSION
(runtime) section — objective + directives — after the immutable
contracts; unreadable path is a hard error, absent env changes nothing
- mosaic-task.mjs: exports MOSAIC_MISSION_FILE as the run snapshot's
container path (/var/lib/mosaic/runs/<id>/mission.json), with an
outside-dataRoot guard; also exports the configured adapter
Verified: contract-only prompt has no mission section; mission-bearing
run shows objective + directives in the generated prompt, snapshot
recorded, real provider returns exactly MOSAIC_HELLO_OK.
Closes#17
- live hello failure dumps latest run result.json + stderr tail before
sandbox cleanup destroys them
- wrong-expectExact case asserts reason == expect-mismatch (was: any
exit 1, which masked compose-level failures)
- repair dangling if/else from the docker-guard refactor
Closes#15
M3 made MOSAIC_IMAGE_TAG required in compose, but run-task.sh never
called load_release — direct task runs failed in compose before any
model call. release.sh paths masked it by exporting the tag to children.
Found by owner-run test-task.sh; failure receipts were in the run
records' stderr.txt.
Closes#14
14 cases: RELEASE validation (valid/invalid/missing), tag consistency,
status on empty state, fault-injected refusal with no pointer + single
valid refusal log line, healthy activation, pointer fields, repeat
activation append-only log, rollback-without-previous refusal.
Harness fix learned the hard way: restore RELEASE from backup inline
after the missing-file case (mv-back restored the mutated file); single
exit trap self-heals the repo state.
Closes#12
- activate: image-presence pre-check + M2 task-runner health gate
(tasks/hello-marker.json exact marker) before atomic pointer replace
(tmp+rename); every attempt appended to activation-log.jsonl
- --fault-injection flips the health expectation to prove the refusal path
- rollback: health-gated re-activation of the previous activated imageTag
from the log; refuses when the image is gone or no previous exists
- status: release, tag, pointer, recent log; safe on empty state
- state lives under <dataRoot>/state/ (config-independent, reset-scoped)
Verified: activate OK; fault-injected refuse with pointer unchanged;
rollback-without-previous refuse.
Closes#11
- RELEASE file: single source of release version (0.0.X until declared stable)
- common.sh load_release(): validates version, derives
MOSAIC_IMAGE_TAG=mosaic-poc-agent:<pi>-r<release> from the pinned pi dep
- compose.yaml: image tag is required env; build/hello/verify call load_release
- verify.sh derives the image name instead of hardcoding it
- package.json version aligned to the same 0.0.X line
Closes#10
- scripts/mosaic-task.mjs: validate | run | list
- Strict v1 schemas: unknown keys rejected; ids/prompt/expectExact/
timeoutSeconds bounds enforced; optional mission file resolved against
the task file and validated too
- run: executes through the config-driven container path with stdin
detached (issue #5 class), SIGKILL timeout (default 120s), trimmed
response capture
- Immutable run records under <dataRoot>/runs/r-<utcstamp>-<rand>/:
task.json + mission.json snapshots (write-once), stderr.txt, result.json
- expectExact gate: mismatch -> status failed, exit 1; result.json is
always written
- scripts/run-task.sh: load_config + bootstrap_runtime_dir before exec
- M2 scope: mission directives are snapshotted for provenance, not yet
injected into the runtime prompt (later policy layer)
Closes#6, closes#7
pi print mode reads piped stdin until EOF; an attached terminal stdin
blocked the one-shot run forever. Automated contexts (closed stdin)
never exposed it. Request text comes from the compose command.
Proven: tail -f /dev/null | scripts/hello.sh now returns MOSAIC_HELLO_OK
in ~4s (previously timed out at 30s); verify.sh remains green.
Closes#5
- verify.sh now calls bootstrap_runtime_dir after load_config; previously a
reset-then-verify flow let Docker auto-create a root-owned mount source
- common.sh: fail with clear guidance when data root exists but is not writable
- README: configuration section, bootstrap usage, selftest entry point
- BUILD-LOG: Phase 5 entries with corrections
E2E (clean slate): 20/20 selftests; bootstrap idempotent; config-driven
hello/verify MOSAIC_HELLO_OK; negative marker exit 1; reset + rerun green;
config checksum unchanged across the entire flow.
Closes#4
jason.woltje