Review finding from scooby. This case does not use run_start, so
install_pane_binaries' symlinks land under a home its launcher never consults
(HOME is the trusted parent here). It resolves mosaic and pi through
MOSAIC_RUNTIME_BIN=$FAKE_BIN instead. Valid path, valid green — and a trap for
anyone who later drops that env var believing the symlinks cover it, which
would break the #1241 binary check rather than exercise it.
Comment only; no behavior change. Harness rc=0.
Refs #1241.
Review finding from scooby: this PR added a failure branch the harness
structurally could not reach. The fake tmux answered `has-session` only for
`=_holder:0.0`, so every non-holder agent landed in the session-is-gone branch
no matter what — the `elif` (tmux still reports the session, no pane PID after
the retries) had zero coverage and no way to get any.
That is the same shape as the bug this PR exists to fix, one layer down: a code
path shipped green where the gate that should measure it cannot. Less severe,
because the branch fails closed at exit 69 rather than reporting success — but
"the harness can't reach it" is the sentence that precedes the next silent
regression, so it gets closed here rather than filed.
`MOSAIC_TEST_HELD_SESSIONS` lets a case name targets the shim should also
answer for. It answers them only AFTER `new-session`, and that detail is the
whole trick: the launcher asks `has-session` about the same name twice — once
at line 255 where a yes means "already running, exit 0", and once at 417 where
a yes means "the session survived". A shim answering yes to both short-circuits
at the first and never reaches the branch under test. It would have looked like
coverage while measuring the idempotency path.
Both failure modes were measured, not reasoned about:
- toggle absent (the old shim): `code=pane-did-not-survive` — the case lands on
the wrong branch, which is exactly the unreachability being reported.
- toggle answering unconditionally: launcher exits 0 via the idempotency
short-circuit — "launcher reported success over a session with no resolvable
pane PID".
- toggle gated on new-session: `code=pane-pid-unresolved`, exit 69.
The case also asserts the diagnostic is not `pane-did-not-survive` and does not
mention the heartbeat, so the two pane faults cannot collapse into one message.
Gates: bash -n · launcher harness rc=0 · test-fleet-units.sh (real tmux) rc=0 ·
fleet specs 342 passed.
Refs #1241.
`mosaic fleet --help` reads "Manage the local Mosaic tmux fleet" and every
roster the CLI scaffolds sets `transport: tmux`, but neither `tools/install.sh`
nor `tools/_scripts/mosaic-doctor` contained the string "tmux" at all. A
greenfield host therefore came out of the installer able to install a fleet,
start a fleet, and run no seat, with `mosaic fleet ps` as the operator's first
and only signal.
Measured on mosaic-sbx-dev (Debian, no tmux, framework installed): `mosaic-doctor`
reported 11 warnings and not one of them named the reason no seat could launch.
The installer gets a warning, not a `require_cmd` hard failure: tmux is required
by the fleet, not by mosaic. Hosts that install this to run `mosaic claude` and
never scaffold a roster are common, and failing their install over a binary they
do not need would be wrong. The check runs in `--check` mode too — "what is the
state of this host" is the question `--check` is asked.
Both checks read the roster's own `transport:` rather than assuming tmux, so a
host declaring something else is pointed at the binary it actually needs instead
of at the wrong package.
The two implementations are deliberately parallel and each carries a comment
pointing at the other. They are separate because the installer must answer this
before the framework's own scripts are guaranteed to be on disk. One harness
drives BOTH from the shipped text — the functions are extracted from the scripts
by awk rather than copied — so the pair cannot drift silently, and the test
cannot keep passing after the shipped copy changes.
The harness is wired into `test:framework-shell`. Without that it would have
tripped the #1017 enumeration guard as UNENUMERATED, which is the guard doing
its job: a check nothing runs is not a check.
Evidence:
- red: the harness fails against origin/next ("could not extract
fleet_declared_transport"); `grep -ci tmux` on both files at origin/next = 0.
- green on real hosts, all four branches:
- dev (no tmux, no roster) -> WARN naming tmux, points at `mosaic fleet init`
- dev (no tmux, v2 roster) -> WARN naming the roster, points at `mosaic fleet start`
- dev installer --check -> WARN saying start "reports success and no seat comes up"
- canary (tmux present) -> `[OK] Fleet transport available: tmux` under --verbose,
silent by default (pass() is verbose-gated), installer silent
- harness green on node:24-alpine/busybox, the CI base image.
- `bash -n` x3, `pnpm typecheck` 45/45, fleet specs 342 passed,
enumeration guard OK, its self-test OK, prettier clean.
Refs #1240. Upstream of #1237/#1243 and #1241/#1244: a correct fix for either of
those still leaves this host with no live seat.
`mosaic fleet start` returned 0 over three dead panes. The launcher knew,
and said the wrong thing at the wrong severity to the wrong layer.
The pane runs `mosaic yolo <runtime>` under PANE_PATH with a cleared
environment. When that binary is absent the pane dies in under a second,
tmux destroys the session, and the diagnostic goes with it. The launcher
then found no PANE_PID, printed a WARNING about the *heartbeat sidecar*,
and exited 0 — so systemd logged "Finished ... successfully" and
`fleet start` reported success. `fleet ps` was the only component telling
the truth.
Two changes, both in start-agent-session.sh:
1. Before any effect, resolve `mosaic` and the roster's runtime against
PANE_PATH — the pane's own view of the path, not the launcher's.
`mosaic yolo <runtime>` calls checkRuntime(runtime) and looks for a
binary named exactly like the runtime, so this asks the same question
the pane will ask a moment later, while an operator can still see the
answer. Absent binary -> exit 69, code=missing-binary, no session
created.
2. Replace the dead-pane WARNING+exit-0. An absent session one second
after new-session is a runtime that died on startup, not a heartbeat
problem -> exit 69, code=pane-did-not-survive, with the command to run
by hand to see why. A present session with no pane PID after five
attempts -> code=pane-pid-unresolved. Neither branch kills the
session; destroying a possibly-live pane on a guess is worse than
leaving it for inspection.
Exit 69 (EX_UNAVAILABLE) is deliberate: the 64s already in this file mean
the projection was bad, and here the data is fine and the host is not
ready. Callers separate the cases by `code=`, the same way fail_env's
codes share 64.
This propagates for free. `fleet start` calls runChecked() for the holder
and each agent, and runChecked throws on non-zero, so layers 4 and 5 stop
lying without a TypeScript change. Two adjacent defects are left for a
follow-up issue rather than widened into this diff: the per-agent loop
aborts on the first failure instead of attempting all and reporting an
aggregate, and runChecked's bare throw surfaces the launcher's message
under a Node unhandled-rejection stack trace because program.parse() is
synchronous.
Tests:
- test-start-agent-session.sh gains three cases: `mosaic` absent from the
pane path, the runtime absent from the pane path, and a pane that does
not survive. Each was verified individually red against the unmodified
origin/next launcher.
- The two cases asserting a valid launch now supply a pane PID. Until now
the suite's one success path was itself a dead pane the launcher
reported as fine.
- The harness fakes `npm` so PANE_PATH stops depending on whatever the
host has installed, and fails loudly if the host provides `mosaic` or
`pi` in the system path, where the missing-binary cases would not be
measurable at all.
- test-fleet-units.sh gains a `pi` shim in its runtime bin. The real-tmux
harness named `pi` in its roster and never installed it; the new
preflight caught it.
Refs #1241
Self-inflicted and worth recording rather than quietly amending.
To prove the new tests were red without the fix I ran
`git checkout origin/next -- <fleet.ts> <[email protected]>`. That writes
the *index*, not just the working tree. Copying my versions back afterwards
restored the working tree only, so the unit file sat staged-as-origin/next and
modified-in-tree, and the next commit (67f5014c) committed the index — silently
removing the ConditionPathExists line that 463745e3 had added.
Nothing caught it. The spec reads the file from the working tree, so it stayed
10/10 green against a HEAD that no longer had the guard. Found by reading
`git status` after the push, not by any gate.
Verified by content, not by assumption:
origin/next 0 occurrences
463745e3 1
67f5014c 0 <- the regression
this commit 1
Refs #1237
Two follow-ups from the canary red->green run and scooby's review.
1. The v2 refusal in `add`/`remove` was a bare `throw`, which reaches the CLI
top level uncaught and prints the guidance under a Node stack trace. The
message *is* the point of the refusal, so it now goes through
`command.error()` — the same clean path the roster-config error uses.
Caught on canary, not in review: the unit tests asserted the message text
and passed either way.
2. The unit-template test asserted only that ConditionPathExists is present.
Presence is not effect. Added two tests for the parts that can drift in
code while that assertion still passes: the condition resolving to exactly
the file the fleet writes (%h/%i rendered against a real install), and the
launcher genuinely failing on an absent generated env (exit 64,
`missing-file`) — which is what makes the condition load-bearing rather
than decorative.
systemd is not available in the suite, so the effect itself was measured on
canary (2026-08-16), roster v2 generation 3:
with the condition: start rc=0, Result=success, ConditionResult=no,
journal "skipped, unmet condition check"
condition removed by
drop-in, nothing else: start rc=1, Result=exit-code, ExecMainStatus=64,
unit failed, "agent environment rejected: missing-file"
Canary red->green for the three commands, same v2 roster, side by side:
fleet ps 0.0.50-next.2413 rc=1 -> branch rc=0 (3 agents listed)
fleet install 0.0.50-next.2413 rc=1 -> branch rc=0
fleet remove <name> 0.0.50-next.2413 rc=1 -> branch rc=1, refusal naming
delete + apply
All three previously failed with "Fleet roster has unknown field(s):
generation." The #791 negative was measured too: the six existing
*.env.generated files were untouched by `install` (mtimes 20+ minutes older
than the run).
Gates: typecheck 0, eslint 0, prettier clean, fleet specs 382 passed, new spec
10/10 with the fix and 9/10 red against origin/next (the 10th passes there for
an unrelated reason and is annotated as such). Full suite: only
mutator-gate.acceptance.spec.ts fails, pre-existing on origin/next.
Still true and still worth saying: a correct fix here shows install rc=0 and
start rc=0 and STILL no live seat. #1240 (tmux absent) is upstream, #1241
(start reports lifecycle-complete over dead panes) and the missing agent
runtime are downstream.
Refs #1237
Reviewed-by: scooby (by git comms; cannot file a Gitea review from fomo-lin)
On a roster-v2 fleet, `ps`, `install`, `install-systemd`, `add` and `remove`
all failed in the v1 parser. The consequence was that a greenfield v2 box could
never get its unit templates placed, so nothing downstream could start.
The read-only commands get a narrow version-agnostic view of the roster
(version, socket name, holder session, and per agent name/alias/runtime).
This is deliberately not a v2 -> v1 downshift. A downshifted FleetRoster would
be accepted by generateAgentEnvValues, which would make a third writer of
fleet/agents/<name>.env.generated through the v1 mapping and break the #791
single-SSOT invariant that projectRosterV2AgentGeneratedEnv is documented to
hold. The view is too small to write a roster or an env file back from, so that
misuse is unavailable rather than merely discouraged.
So on a v2 roster `install` places the tool files and the unit templates,
enables the units, and writes no generated env at all. Env belongs to `apply`
and `regen`, both already v2-native.
That change alone would have traded an init-time failure for a boot-time one.
`install` enables mosaic-agent@<name>.service (WantedBy=default.target) without
starting it, so a reboot between `install` and the first `apply` would run
ExecStart against an absent env file and fail every seat unit, further from its
cause. The unit template now carries
ConditionPathExists=%h/.config/mosaic/fleet/agents/%i.env.generated
which skips an enabled-but-unconfigured unit cleanly and starts it on the next
start once the reconciler has written env. On v1 it is a no-op, since v1
`install` writes env itself. Found in review by scooby.
`add` and `remove` are not routed to `create` and `delete`. They are different
operations: the v1 pair edits the roster and drives systemd, the v2 pair is
documented as changing desired state without runtime actions. `add` also
collects four fields where a v2 agent requires eleven, so routing it would mean
inventing an operator's provider, alias, reasoning and tool policy. On v2 both
now fail with the real two-step sequence instead.
Tests: 8 new, 7 of which are red before this change. Includes the greenfield
case scooby asked for — `ps` on a fresh v2 install with nothing running is rc=0
and lists every agent stopped, since that is the command an operator runs to
find out why there is no seat.
Note for anyone verifying this: a correct fix here shows `install` rc=0 and
`start` rc=0 and still no live seat. #1240 (tmux absent) is upstream, #1241
(start reports lifecycle-complete over dead panes) and the missing agent
runtime are downstream. A dead pane after this change is not a regression here.
Refs #1237, #791, #1240, #1241
scooby's review flag 1 on #1242. The other three chmods warn; this one was
`|| true`. It is the one directory holding secrets, so a chmod that fails
silently there is the failure most worth a line in the output.
Comment-and-warn only. No behaviour change on the success path.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01WYgWocp36goy8hj2ui6ps1
A greenfield install cannot run `mosaic fleet init --write`. It fails with
`unsafe-permissions` on an unnamed `(directory)` and an unhandled Node throw,
and every mutating `mosaic fleet` command fails the same way. Measured on a
reverted-to-greenfield sandbox VM at CLI 0.0.50-next.2413: `~/.config/mosaic`,
`fleet/` and `credentials/` all land at 0775, and 1735 directories under the
framework root carry `mode & 022`.
Two independent causes, and fixing either one alone leaves it broken.
1. The installer inherited the caller's umask. Debian/Ubuntu ship 002, so every
`mkdir -p` produced 0775. Fedora/RHEL ship 022 and produced 0755. The
product therefore worked or did not depending on the operator's login shell,
with nothing in the install output distinguishing the two. 022 is already
what this script assumes it produces — `make_durable_snapshot` restores the
ambient umask specifically so "every later sync copy and new framework dir"
gets 0644/0755 — so pin it rather than inherit it.
2. Even at a correct 0755, three directories are rejected. The fleet code
guards its managed paths with two masks in two languages:
`assertPrivateManagedDirectory` (fleet-reconciler.js, `mode & 0o077`) covers
MOSAIC_HOME and `fleet/` and runs before the roster lock is taken;
`assert_private_directory` (tools/fleet/start-agent-session.sh, `mode & 077`)
covers `fleet/agents` and runs before a pane is spawned. Their laxer
siblings use `mode & 0o022` and accept 0755. The strict mask wins, so the
installer states 0700 outright instead of hoping a umask implies it.
The `find -perm /022 -exec chmod go-w` sweep repairs a tree installed before
this change, which the umask alone cannot reach. It strips group/other WRITE
only — never read or execute — and is scoped to directories, so it corrects the
boundary violation without changing who may traverse or read anything. It is
not sufficient for `fleet/agents`: stripping write from 0755 yields 0750 and
`mode & 077` is still non-zero, which is why that path gets its own chmod.
Reported as #1236. The `fleet/agents` half was found by scooby reading
start-agent-session.sh; the umask framing is theirs too — my first report
blamed the distro rather than the umask.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01WYgWocp36goy8hj2ui6ps1
getShellProfilePath() preferred ~/.bashrc when it existed, and ~/.zshrc for
zsh. setupPath() in stages/finalize.ts appends the PATH export to whatever
it returns. Debian's default ~/.bashrc opens with
case $- in *i*) ;; *) return;; esac
so a line appended to the bottom of it never runs for 'bash -lc', for
systemd units, for 'ssh host cmd', or for any agent seat — precisely the
consumers that need the CLI. An install could print its summary and exit 0
while leaving 'mosaic: command not found'. .zshrc has the same problem:
zsh only reads it for interactive shells.
Now ~/.profile, which login shells read and which Debian's copy sources
.bashrc from for interactive shells, so one line covers both. For zsh the
always-sourced file is .zshenv. fish and PowerShell are unchanged.
__tests__/platform/detect.test.ts pins it, including a case asserting that
no shell resolves to an interactive-only rc file. Falsified by inverting
the fix: 5 failed / 1 passed; restored 6/6. Full package suite unchanged at
17 files / 4 tests failing, matching clean origin/next.
runRuntimeLaunchEntry set MOSAIC_HOME to the shipped framework root, so
launch-runtime.py appended its launch ledger to
framework/fleet/run/sessions/events.ndjson — polluting the tree that
manifest.spec.ts walks and failing its completeness check in CI.
Point MOSAIC_HOME at the per-entry temp root instead. Nothing in the
launch chain resolves tools via MOSAIC_HOME (entry scripts resolve via
SCRIPT_DIR); the ledger is its only consumer here.
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Dtdjx4Gxude9fwyLezCrhh
The two install-linker-compatibility tests still asserted the pre-isolation
behavior (mosaic skill links planted in $HOME/.claude/skills). This branch
deliberately moved the link farm into the mosaic-owned harness homes
($MOSAIC_HOME/.claude/skills) and demoted the base-install dirs to
cleanup-only legacy targets, so the tests now assert the new topology:
the skill links appear under the harness home, foreign links in the legacy
dir are preserved, and no new mosaic link is planted in the base install.
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Dtdjx4Gxude9fwyLezCrhh
The mosaic wrapper makes pane_current_command=node (RUNTIME_ACCEPTABLE_COMMANDS.claude=['claude','node']); the walk matters precisely in that no-shell-wrapper case. Match reality.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
Claude-Session: https://claude.ai/code/session_013SAYFkRhQfhguY7AHfiUC8
The launcher runs the runtime as a spawnSync CHILD of node(mosaic) (deliberate,
per launch.ts:99 — parent survives to propagate signals), so
MOSAIC_LEASE_SESSION_ID lives on the claude child, not the pane's root pid. The
transport read only pane.pid's /proc/environ and returned RESOLVE_FAILED for
every real 'mosaic claude' seat. Now BFS the pane's process subtree (bounded,
injectable children-reader) and read the first descendant that carries a valid
lease id; fail-closed if none. Unit tests now exercise the real walk (pane=node
without lease -> child=claude with lease) rather than mocking the resolution.
Found by scooby greenfield E2E on fomo-lin with proc-level evidence.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
Claude-Session: https://claude.ai/code/session_013SAYFkRhQfhguY7AHfiUC8
MAX_LEASE_TTL_SECONDS (daemon cap+default) and DEFAULT_TTL_SECONDS
(lease_promote client) both move to 3600. The 5-minute TTL made
gated-by-default sessions unusable (re-promotion mid-task); 1 hour
matches a working session. Full test:framework-shell RC=0.
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_013SAYFkRhQfhguY7AHfiUC8
The independent W-0R review of 3592b92e passed but left two PLAUSIBLE
findings: the stderr notice for a legitimately-omitted operator source was
claimed and never asserted (a silent omission is the original defect in
miniature), and the chmod 0o000 unreadable simulations fail spuriously when
euid==0 (CAP_DAC_OVERRIDE). Falsifier for the new assertion: deleting the
notice block turns the suite red (failures=3); restoring returns green.
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EHYXhcCQsL3J1Lnm7EraGq
build_construction skipped any normative source it could not read
(`except OSError: continue`) and promoted whatever remained. That is not a
degraded binding, it is a forged smaller one: the broker recomputes h_source /
h_payload from the fragments it is SENT (daemon.py:602-616), so an omitted
fragment is internally consistent and PAYLOAD_BINDING_MISMATCH cannot fire. A
partial law promotes exactly like a complete one and nothing downstream can tell
the difference.
Measured before this change, against a seeded home: with only USER.md readable,
the client produced a one-fragment construction with promotion=True. Removing
CONSTITUTION.md, STANDARDS.md or the runtime contract likewise promoted.
The classification mirrors the framework's own file ownership rather than
inventing one:
* CONSTITUTION.md / AGENTS.md / STANDARDS.md are framework-owned and
reconciled every upgrade (install.sh FRAMEWORK_OWNED,
config/file-adapter.ts FRAMEWORK_OWNED_FILES), as is the per-runtime
RUNTIME.md. Absent => IncompleteBinding. A deployment missing one is broken,
not minimal.
* SOUL.md / USER.md are deliberately not seeded by install.sh ("generated by
`mosaic init`") and TOOLS.md is seeded on first install only, so their
absence is legitimate. It is reported on stderr, never silent.
Unreadable is handled separately from absent for EVERY source, optional ones
included: a file that will not open is not a file that was never configured, and
collapsing the two is what let a permission change quietly shrink the law.
Also corrects this module's own docstring, which asserted that a VERIFIED lease
means "this agent is running THIS law". It does not. Both sides of the broker's
comparison originate in this client, so it detects corruption in transit and
nothing else. That overstatement is where the belief spread from; the stronger
claim needs the broker re-reading on-disk sources against a manifest the agent
cannot rewrite.
Test: promotion_binding_unittest.py, enumerated in test:framework-shell (the
enumeration guard's population is *test*.sh and does not cover Python, so an
unenumerated test here would simply never run). Falsifier executed: defeating the
guard while leaving the module API intact turns the suite red (12 failures);
restoring it returns green.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EHYXhcCQsL3J1Lnm7EraGq
This reverts 939f2e04. Keeping the revert rather than dropping the commit,
because the failed attempt is the most useful record on this branch.
The wiring worked mechanically — verified with a live model on sb-it-1-dt: the
receipt was emitted verbatim as a whole message, and the broker token was minted
AND consumed, so observe_receipt and promote_lease both succeeded and the lease
reached VERIFIED.
It failed as a DESIGN, for reasons that are properties of the protocol rather
than of this wiring:
* It puts control-plane traffic in the user-facing conversation channel. An
operator asking "what model are you?" received a receipt string instead of an
answer — the model tried a tool, was blocked, complied with the receipt
instruction, and in one-shot mode that text turn BECAME the reply. Observed
twice, non-deterministically.
* The lease TTL is hard-capped at 300s (MAX_LEASE_TTL_SECONDS; ttl_seconds >
cap raises INVALID_LEASE_TTL). Measured: allowed at T+0, LEASE_EXPIRED at
T+310. So the visible cost recurs every five minutes of mutator activity.
* Model compliance is not guaranteed — one run retried the command instead of
emitting the receipt.
Any model emission is user-visible, so this is not fixable by better wiring; it
needs a design answer about how promotion is triggered and paid for. That is
under adversarial review (docs/scratchpads/lease-remediation/07-liveness-design-brief.md
in the operator's repo). Promotion triggering will return on its own branch once
that lands.
What remains here is independently sound and unblocked: harness-home isolation,
the immutable launch record, the skills relocation, the promotion client itself
(steps 1/4/5), and the #1087 prefix guard.
prune_stale_links_in_target compared "$resolved" == "$canonical_real/"* while
length-checking only $resolved. If $canonical_real were ever empty the pattern
collapses to == "/"* and matches every absolute path.
The failure is precisely inverted, which is what makes it worth fixing rather
than noting: is_mosaic_skill_name already `continue`s for names that ARE current
mosaic skills, so an empty prefix would delete exactly the FOREIGN symlinks in
every target directory and preserve the mosaic ones. On this host that is 4 base
installs, including codex's own .system entry.
Reported by mos-claude as #1087 after I introduced the same guard in the new
legacy-cleanup path in the previous commit and walked past this instance thirty
lines away. Same defect class, same file, one function apart.
$canonical_real is populated by readlink -f after a mkdir -p, so an empty value
requires readlink to fail — unlikely, but the consequence is deleting operator
symlinks across every harness, which is not a risk worth carrying for one test.
Completes the promotion path: the client landed in the previous commit, but
nothing drove step 2 — the model emitting the receipt. This wires it.
LAZY, not at session start. Promotion costs an entire model turn, because the
receipt must be the whole message (hmac.compare_digest, "not a transcript
substring"). Minting at session start would collide with the Constitution's
first-response mode declaration — the two cannot share a message, so requiring
both would be unsatisfiable. Deferring to the first DENIED MUTATOR means the
mode declaration happens first and the receipt gets its own later turn, so no
governance change is needed. A read-only session never pays for promotion at all.
Mechanism: on a MUTATOR_UNVERIFIED denial the tool_call hook mints a challenge
and returns the receipt in the block `reason`, which pi feeds back to the model
as the tool result — the existing injection path already used by
lease-lifecycle.ts. The model emits the receipt as its next message, message_end
ships it to the observer, and the extension then calls observe_receipt +
promote_lease.
Only MUTATOR_UNVERIFIED triggers minting. Other denials (GATE_UNAVAILABLE,
STALE_GENERATION, LEASE_EXPIRED, ANCESTRY_MISMATCH) describe conditions a
receipt cannot fix, and begin_verification revokes before it mints, so minting
there would thrash the broker.
Completion is gated on an EXACT text match against the minted receipt. This is
load-bearing, not defensive: message_end also fires for the message that
CONTAINED the blocked tool call — one turn BEFORE the model answers. An earlier
version completed there, so observe_receipt compared against the wrong text,
failed, and burned the challenge before the model ever emitted it. Matching the
text mirrors the broker's own compare_digest semantics and waits for the right
turn. Confirmed by instrumenting message_end and watching it fire with
pending=yes one message too early.
It never posts the receipt itself. receipt-observer-client.py accepts any
string, so self-posting would satisfy the broker while proving nothing — the
whole point is that a live model echoes a challenge it was given.
Bounded by MAX_PROMOTION_ATTEMPTS: model compliance is not guaranteed (observed
a run where the model retried the command instead of emitting the receipt), so
a non-complying model degrades to today's behaviour — denied mutators — rather
than looping.
Verified with a live model on sb-it-1-dt: receipt emitted verbatim as a whole
message, and the broker token was minted AND consumed, i.e. observe_receipt and
promote_lease both succeeded and the lease reached VERIFIED.
Known limitation: under `pi -p`, the receipt is a text-only turn, which ends the
one-shot loop — so promotion completes but the blocked tool is not retried in
that same invocation. Interactive and durable fleet sessions continue and retry
normally.
The enforcement half of the lease broker ships and denies; the promotion half
has no production caller anywhere in the package. Verified across 0.0.48, 0.0.49
and 0.0.50-next.2207: begin_verification / observe_receipt / promote_lease are
invoked only by broker-test-client.ts, the acceptance spec, unit tests, and two
probes under docs/.
Consequence: no lease on any host can reach VERIFIED, so mutator-gate denies
every mutator with MUTATOR_UNVERIFIED via a gate that nothing shipped can
satisfy. Runtimes that enforce the gate in-process (pi, via mosaic-extension's
tool_call hook) are bricked for mutators; runtimes whose gate is wired through a
settings hook escape only when that hook is absent — i.e. by being ungated.
This adds the client. It implements protocol steps 1, 4 and 5:
1. begin_verification -> mint a challenge, return the exact receipt text
2. the MODEL emits that text verbatim as its entire latest message
3. the runtime adapter ships that message to the observer socket
4. observe_receipt -> PENDING_PROMOTION
5. promote_lease -> VERIFIED
Step 2 is deliberately NOT implemented here, and that is the point.
is_verbatim_receipt uses hmac.compare_digest against the exact minted string —
explicitly "not a transcript substring" — which makes promotion a LIVENESS
PROOF: it requires a live model that received the challenge in its context and
echoed it exactly.
receipt-observer-client.py will post ANY string as the latest assistant message.
A promotion client that posted its own receipt would satisfy the broker while
proving nothing — a gate-disabler indistinguishable from a working fix unless
someone specifically looks. Emitting the receipt therefore belongs to the runtime
adapter, where a real model turn happens. A local diagnostic that posts its own
receipt exists in the operator's repo and is deliberately NOT shipped here.
The construction binds the exact normative source bytes, so a VERIFIED lease
means "this agent is running THIS law", not merely "this session id is known".
h_source/h_payload are derived by importing the framework's own
normative_fragments.build_payload rather than reimplementing it: the broker
derives them the same way and any divergence yields PAYLOAD_BINDING_MISMATCH.
There must be exactly one implementation.
session_identity() prefers the generation FILE over the env var, matching
lease_generation.py. Sending a generation higher than the broker's would revoke
the session's own authority (daemon.py:342-344), so it never guesses.
Verified end-to-end on sb-it-1-dt under a real lease-gated anchor: a mutator
denied rc=2 MUTATOR_UNVERIFIED, then begin -> observe -> promote -> VERIFIED,
then the same mutator allowed rc=0. Negative controls pass: a fresh session is
still denied, and an unrelated session still reads UNVERIFIED — promotion is
per-session and does not leak.
Still open: adapter wiring for step 2. Lazy promotion on first mutator attempt
avoids colliding with the Constitution's first-response mode declaration, since
compare_digest requires the receipt to be the WHOLE message.
Mosaic wrote into the operator's harness base installs — ~/.claude,
~/.pi/agent, ~/.codex, ~/.config/opencode — for settings, instructions, and a
102-symlink skill farm per harness. Any experiment with hooks or gating
therefore mutated the operator's own tooling, and a broken framework change
could take out the very harness needed to repair it.
Harness home isolation
----------------------
Each runtime now reads config from a dedicated mosaic-owned home via the
harness's own config-dir variable:
claude CLAUDE_CONFIG_DIR ~/.config/mosaic/.claude
pi PI_CODING_AGENT_DIR ~/.config/mosaic/.pi (replaces ~/.pi/agent)
codex CODEX_HOME ~/.config/mosaic/.codex
opencode XDG_CONFIG_HOME ~/.config/mosaic/.opencode
These paths are manifest-UNKNOWN, so rule 3 (#791) resolves them to operator
ownership and a keep-mode upgrade can neither overwrite nor prune them.
A bare `claude` / `pi` keeps its own config AND auth, making it a structural
break-glass rather than one depending on restoring a file under pressure.
opencode is blunter than the rest: it has no dedicated variable and follows XDG,
so isolation also relocates XDG lookups for anything it spawns. Documented in
place.
mosaic-sync-skills now links into those homes and cleans the legacy farms it
previously planted in base installs. Ownership is proven by RESOLUTION, not by
name — only symlinks resolving inside the canonical/local skills dirs are
removed, mirroring the refusal already in commands/skill.js. Verified against a
real install: codex's own .system directory survived while its 102 mosaic links
were removed. Both resolution prefixes are length-checked first; an empty prefix
would make "$resolved" == "$prefix/"* match every absolute path and delete
foreign symlinks.
Immutable launch record
-----------------------
Every launch now appends one record to fleet/run/sessions/events.ndjson before
exec. Mandatory, mechanical, no model involvement.
pi rewrites its own argv to a bare `pi`, so /proc/<pid>/cmdline destroys the
launch evidence — that has already produced a confident wrong diagnosis ("this
agent bypassed the launcher"), disproved only by the parent's argv and only
because the parent had not yet exited. A record written before exec is the only
place this survives.
The path is the #797 Runtime Session Ledger, already operator-classified and
already covered by test-upgrade-manifest-guard.sh, which seeds it and proves a
populated ledger survives keep-mode upgrades — but nothing shipped ever wrote
it. This implements it in the shape that guard already asserts (0600 files under
a 0700 dir).
`mosaic` writes session.launch; launch-runtime.py appends lease.register with
the broker session id and activation capability. They correlate by an explicit
MOSAIC_LAUNCH_ID, never by pid: execRuntime uses spawnSync, so the runtime is a
child with a different pid.
Records normative fragment digests (CONSTITUTION/AGENTS/SOUL/USER/STANDARDS/
TOOLS/RUNTIME) — the same set the broker hashes for promotion, so drift is
mechanically detectable rather than a matter of judgement.
Credential-safe: env is captured as PRESENT NAMES ONLY, and argv values over
256 bytes become a sha256 + length rather than being inlined.
Also fixes CLI_VERSION resolution: '@mosaicstack/mosaic/package.json' is not in
the package exports map and always throws ERR_PACKAGE_PATH_NOT_EXPORTED.
resolveTool() uses that same failing specifier, which is why its documented
preference for bundled tools over the deployed ~/.config/mosaic copy has never
once applied — noted in place, not fixed here.
Verified on sb-it-1-dt: isolated homes written and base installs byte-identical
for all four harnesses; 408 legacy symlinks removed with 1 foreign entry
preserved; launch records paired across the spawn boundary. typecheck shows zero
errors in launch.ts (the @mosaicstack/types failures are pre-existing and
reproduce on a pristine origin/main worktree).
The `format` step of .woodpecker/ci.yml:89 (`pnpm format:check`) failed on
pipeline 2111 for this branch. Reproduced on a bench with the lockfile-pinned
[email protected] against the repo .prettierrc and .prettierignore, using CI's
exact glob: WRITING-STYLE.md was the only failing file.
The change is mechanical and semantically null: markdown table cell padding
and `*emphasis*` -> `_emphasis_`. Verified by normalizing both revisions
(whitespace removed, `_`/`*` folded, table rules collapsed) — the results are
byte-identical.
This does not address the prose findings published on #965 (P1-P4); those
await a ruling. The `test` step also failed on 2111, on a base ~40 commits
stale — attribution for that failure needs this rerun, and is not claimed here.
Co-authored-by: mos-dt-0 <[email protected]>
§5 sent the agent to read direct|friendly|formal in USER.md, but the builder
renders prose bullets, not the token — the documented lookup could not key on
the shipped file. Table now keys on the leading bullet USER.md actually
contains. Also: 'concise, technical' -> 'concise, structured' (drop the round-1
residual value name from a rule-9 guide). Docs-only, no code, no scope growth.
Written-by: jarvis (dragon-lin)
Co-Authored-By: Claude Fable 5 <[email protected]>
F1: defaults/USER.md is never installed (generated from templates/USER.md.template
via buildCommunicationPrefs). Editing it was a no-op asserting a phantom setting —
exactly the false-green §2 warns against. Reverted.
F2: the framework already has communicationStyle (direct|friendly|formal). §5 now
maps THOSE values to output instead of inventing technical|prose|brief (rule 9).
Minor: §6 states no mechanical prose check exists today; rule 1 points at §3.4.
Written-by: jarvis (dragon-lin)
Co-Authored-By: Claude Fable 5 <[email protected]>
Adds the agent output standard to the framework SOT so it injects at launch and
is selectable per user (closes the gap: it lived only as a jarvis-brain lab doc + issue #960).
- guides/WRITING-STYLE.md: MOS-STE (adapted ASD-STE100) for docs, Google Style for code,
verification-artifact emphasis, absolute user-voice carve-out. Written in MOS-STE.
- defaults/STANDARDS.md: Output-standards block (always injected via the prompting contract).
- defaults/AGENTS.md: routing row so writing/doc/comms work reaches the guide.
- defaults/USER.md: per-user 'Comms style' option (technical|prose|brief), default technical.
Refs mosaicstack/stack#960. Owner directive (Jason, 2026-07-30): docs->adapted ASD-STE100,
code->Google style, resumes/personal carved out, comms style a per-user choice.
Written-by: jarvis (dragon-lin)
Co-Authored-By: Claude Fable 5 <[email protected]>
fred
mos-dt-0
shaggyandmos-dt-0
be-coder-08andJason Woltje
