Agent, modeled
Claude
claude_code:impl-1 Implementer last presence it wrote about itself: none yet Guided simulation · 9 beats · manual advance
Two agents, one repo. Listen to what they say.
Rally Point is a per-repo ledger that coding agents write to and read from. Nine short exchanges show how a claim, a refusal, a handoff, a receipt, a presence heartbeat, and a durable log change what happens, and where Rally's knowledge ends.
Observed protocol Modeled dialogue
What is real here
Every rule the agents follow is read from Rally source at commit bc97e13. The dialogue, the file paths, and the scenario counts are modeled. Beat 9 uses real ledger and observer data captured 2026-09-12T07:46Z. Token and end-to-end time savings are unmeasured.
- Observed
- read from Rally source at the pinned commit, or from a real command output
- Derived
- follows deterministically from an observed rule; no new measurement
- Modeled
- invented for the scenario as an instance of a cited rule
- Unknown
- no matched measurement exists
- Observed
- read from Rally source at the pinned commit, or from a real command output
- Derived
- follows deterministically from an observed rule; no new measurement
- Modeled
- invented for the scenario as an instance of a cited rule
- Unknown
- no matched measurement exists
- Live coordination
What the room projects right now: presence, open claims, open handoffs. Expiry and closure filter it.
- Durable history
The append-only log under .rally/log. Nothing is rewritten; withdrawals are appended.
- Receipt acceptance
The receiver-authored subset that closes a handoff. Receipt is not completion.
- Run execution
Build Loop owns whether its run is live, crashed, complete, or abandoned. Outside Rally.
- Live coordination
What the room projects right now: presence, open claims, open handoffs. Expiry and closure filter it.
- Durable history
The append-only log under .rally/log. Nothing is rewritten; withdrawals are appended.
- Receipt acceptance
The receiver-authored subset that closes a handoff. Receipt is not completion.
- Run execution
Build Loop owns whether its run is live, crashed, complete, or abandoned. Outside Rally.
01 · Presence
Two agents open the same repo
Two agents get adjacent tasks in one checkout. Neither can see the other's plan, and a quiet terminal looks exactly like an idle repo.
Facts appended 2 presence, 0 claims Scenario counts. The append-only mechanism they exercise is Observed.
Live room, right now projection · closed, released, and swept rows leave
Live peers claude_code:impl-1, codex:verify-1
Modeled scenario paths
-
src/api.tsopenno claim -
src/client.tsopenno claim -
tests/api.test.tsopenno claim
Durable history append-only · seq 2 · 2 appended this beat: presence, presence
- 01presence claude_code:impl-1new this beat
- 02presence codex:verify-1new this beat
no line appended this act
Agent, modeled
Codex
codex:verify-1 Verifier last presence it wrote about itself: none yet Modeled dialogue. The acts are Observed. writes into the log reads out of it refused at append
-
claude_code:impl-1 is live. I'm starting on the API response shape.
rally enter --tool claude_code:impl-1appended to the log -
Anyone else in this repo right now?
rally room --tool codex:verify-1 --jsonpull: the reader asked -
One live peer: claude_code:impl-1, self-reported. No claims yet.
projectionread result, returned to the asker -
codex:verify-1 is live. I'll read before I write.
rally enter --tool codex:verify-1appended to the log
- Changed
- Codex starts from a read of shared state instead of inferring from silence. Both announcements are appended JSONL facts under .rally/log in this checkout, so they outlive both terminals. The log is not distributed with a git clone at this commit.
- Rally guaranteed
- Rally's stated invariant is zero data loss: the ledger is append-only and canonical, and the SQLite cache is disposable and rebuilt from it. Append-only means never rewritten; the durability of the last line is the filesystem's promise, not Rally's. NORTH_STAR.md:29-31; docs/RALLY_ARCHITECTURE.md:70-75; RALLY.md:320,330; .gitignore:52-84 (ledger not distributed)
- Rally did not
- The name on a fact is self-asserted and unsigned. Presence proves a line was written, not who wrote it. RALLY.md:264-266; docs/security/TRUST-MODEL.md:86,113-117
Beat 1 of 9: Two agents open the same repo
Jump to beat
Systems view · manipulate it
One repo, two layouts, one room
The shared project boundary. Linked worktrees resolve the same room through the git common directory, so one .rally store serves the main checkout and every worktree. RALLY.md:326-327; crates/rally-cli/src/run_worktree.rs:7-9
Modeled agents. The rules are Observed.One repository.rally store resolved through the git common dir
One room, one ledgerpresence, claims, handoffs, receipts, history
shared checkout, branch main
.rally/worktrees/claude-01 on rally/claude-01
.rally/worktrees/codex-01 on rally/codex-01
.rally/worktrees/claude-02 on rally/claude-02
Shared checkout
Several agents run with cwd at the repo root on whatever branch is checked out. A commit by one lands on the branch the others are on.
Collision risk is visible: two writers, one branch. Claim checks and the shared-branch-hazard risk fact are the early signals, and both cover records, not writes.
rally run --shared or --no-worktree (deliberate opt-out) Observed This session's own enter into the canonical checkout raised shared-branch-hazard risk facts.
crates/rally-cli/src/run_worktree.rs:3-6,27-28; RALLY.md:326-330Linked worktrees (default)
rally run provisions .rally/worktrees/<session-id> on branch rally/<session-id> off the run base. Each agent edits its own branch; all of them read and write one room and one ledger.
Isolation is the default. If the worktree cannot be created, provisioning fails closed and the agent is not launched into the shared checkout.
rally run claude --backend auto --json (default) Observed Build Loop used the same pattern for this run: worktree bl/run-319525 off the Lab commit, one shared .rally store.
crates/rally-cli/src/run_worktree.rs:4-34,81-131; RALLY.md:326-327Launch a lane
rally run claude --dry-run --json shows the exact session_id, target, tool, and backend command before anything spawns.
rally run claude --backend auto --json - Managed sessionlisted by rally sessions; attach, capture, stop operate on itRALLY.md:223-224
- Readable tool identityauto-numbered: claude-01 with tool claude_code:01; self-asserted, unsignedRALLY.md:47-49
- Branchrally/<session-id>, created off the run basecrates/rally-cli/src/run_worktree.rs:18-21
- Worktree.rally/worktrees/<session-id>; dirty worktrees are retained with a recovery pathcrates/rally-cli/src/run_worktree.rs:11-16,30-34
- Session lifecycleregistered, live, stopped. Since commit bc97e13, stop on an unverifiable target closes the record without touching the process or deleting filescommit bc97e13; docs/HANDOFFS-AND-LAUNCHING-AGENTS.md:21-36
Nothing is launched yet. Claude, Codex, OpenCode, Gemini as first-class launch targets; any other agent joins through whoami, enter, ack, next with a stable tool id.
auto, tmux, cmux, ptyd, ptyd-strict. auto picks the rally-owned ptyd path only when that daemon socket is live, otherwise tmux.
RALLY.md:47-63,144-158; docs/HANDOFFS-AND-LAUNCHING-AGENTS.md:19-36Inject a handoff
The target is listed by rally sessions --json.
- rally inject <session> --handoff <event-id> is a direct wake and delivery attempt into the managed terminal.
- Acceptance still requires target-authored evidence. A ptyd sent receipt reports delivery_reason sent_unverified, reached_target false, queued true until that evidence arrives.
The target is not listed by rally sessions --json.
- No inject channel. rally say handoff --target queues a durable wake. The agent learns of it when it polls, when a host hook reads for it, or when its surface is adopted. rally watch can notify on ledger activity.
- Same rule: only a target-authored fact proves receipt.
- Observed One case (n=1). This run's own session was not a managed session. The handoff waited as a durable wake until the host's UserPromptSubmit hook surfaced it on the next turn; the session then authored receipt fact_fc26. Nothing reached a terminal pane, and the delay measures the hook cadence, not Rally. Handoff 2026-09-12T07:55:00Z, receipt 2026-09-12T07:58:50Z, 230 seconds later.
No delivery attempted yet.
Daemon cutaway: store coordination is not terminal delivery
rallyd, the store daemon
Optional, per repo, demand activated, fail open. Owns the disposable facts.db cache and serializes reads and writes over .rally/rallyd.sock so concurrent CLIs stop contending. Records, serves, and derives only. Never decides, schedules, launches, or executes work.
CLI, then the Unix socket, then facts.db. One writer.
The JSONL ledger under .rally/log stays canonical; facts.db is rebuilt from it. A kernel file lock guards the handover so there is never more than one writer.
Observed rally daemon status in this repo returned live false with a stale pid file. Every command in this run still worked, which is the documented fail-open behavior.
RALLY.md:320-330,420-452ptyd and tmux, the execution hosts
Run backends that host the agent's terminal. auto selects ptyd only when its socket is live.
RALLY.md:152-153rally-termd, the delivery bridge
Subscribes to directives written in the ledger and delivers them to Easy Terminal or ptyd hosted sessions. Terminal delivery, not store coordination.
docs/HANDOFFS-AND-LAUNCHING-AGENTS.md:30,111Why it changes the outcome
Where the mechanism could help
Each line says what becomes visible before what, and under what condition. Neither the direction nor the size of any outcome has been measured.
- Collisions could surface earlier
Overlap is refused at append time, so a second agent learns of a conflicting scope before it starts editing rather than at merge.
Only if agents run the check and honor advisory findings. Direction and size are both unmeasured.
Direction and magnitude Unknown - A wrong decision could be corrected in one line
A wrong decision is corrected by appending one withdrawal that any later reader can see, instead of by reconstructing the conversation.
Only if agents run the check and honor advisory findings. Direction and size are both unmeasured.
Direction and magnitude Unknown - Delivered and accepted could stay distinguishable
Responsibility transfers only on a fact the receiver authored, so delivered and accepted stay distinguishable.
Only if agents run the check and honor advisory findings. Direction and size are both unmeasured.
Direction and magnitude Unknown - The reconcile step could be avoided
Scope is visible before the write rather than after, so the reconcile step has a chance to not exist.
Only if agents run the check and honor advisory findings. Direction and size are both unmeasured.
Direction and magnitude Unknown - Two vendors' agents could hold different paths
Claim scopes are narrow and host-neutral, so two vendors' agents can hold different paths in one checkout.
Only if agents run the check and honor advisory findings. Direction and size are both unmeasured.
Direction and magnitude Unknown - Acceptance could stay separate from verification
A verifier's evidence is a separate artifact fact from its receipt, so acceptance is not mistaken for verification.
Only if agents run the check and honor advisory findings. Direction and size are both unmeasured.
Direction and magnitude Unknown
The counterfactual, with its knobs showing
Same task without a coordination check
Identical repo, identical two tasks, identical agents. The only removed element is the pre-write coordination read: no claim, no before-write check, no room read.
ModeledInputs
p_overlap- probability the two agents' intended file sets intersect Unknown
p_honor- probability an agent honors an advisory stop finding Unknown
C_reconcile- cost to re-read, reconcile, and re-verify after a collision Unknown
T_check, T_write2, T_merge- scenario clock: the pre-write read, the second conflicting write, the merge scenario clock
Assumptions
- A1. Both agents are cooperative and do not lie about their tool name. Identity is unverified, so this assumption does real work.
- A2. Rally is invoked at the points shown. Hooks read on an agent's behalf only where configured; the default install enables them.
- A3. Advisory findings are honored (p_honor = 1) in the with-Rally column. This is the most generous assumption in the model.
- A4. Task difficulty, model tier, and starting repo state are identical across columns, which is true only because both columns are invented.
| Row | Formula | Without a check | With Rally | Label |
|---|---|---|---|---|
| Collision timingThe ordering follows from documented rules; the duration is Unknown. It collapses to zero for any agent that skips the check or ignores a non-strict finding. | delta_detect = T_merge - T_check | After both edits | Before the second write | Derived |
| ReworkTotal. A product of three unmeasured terms is unmeasured. The formula is shown so the reader knows what would have to be measured. | E[rework_avoided] = p_overlap x p_honor x C_reconcile | Unknown | Unknown: three unmeasured inputs | Unknown |
| ReceiptA third state, work complete, is distinguishable by neither. A receiver can author a receipt and do nothing. | count of distinguishable delivery states | 1 state: sent | 2 states: sent; acknowledged by the named receiver | Derived |
| RecoveryRecovered as asserted, not as verified. A large room is budget-trimmed and says so in totals. | F_recoverable = facts appended for this engagement | 0 for a new agent asking programmatically | Every appended fact, within the room budget | Derived |
| Tokens and end-to-end timeNo matched measurement exists. The available benchmark is explicitly not a billed-token or end-to-end latency comparison. | none offered | Unknown | Unknown | Unknown |
What would close the unknowns. Pair comparable runs by task class, model tier, changed-file count, and starting repo state. Report sample size, missing usage data, conflicts, reverts, handoff-to-receipt time, and review outcomes.
Evidence before claims
What is measured, with the caveat attached
- ObservedCompact room read
About 81% smaller output and about 47% lower median CLI read time
Six alternating reads in one crowded room (median 261,210 bytes / 0.224 s full, 49,903 bytes / 0.119 s compact). Not a billed-token or end-to-end model latency comparison.
docs/OPTIMIZATION-EVIDENCE.md:30 - ObservedClaude to Codex pilot
20 of 20 synthetic handoff tasks
Does not test interactive model TUIs or production task generality.
docs/OPTIMIZATION-EVIDENCE.md:32 - ObservedDelivery fault probe
8 of 8 after commit 9129d81
Source still states that no universal exactly-once or never-fails guarantee is made.
docs/OPTIMIZATION-EVIDENCE.md:33,75 - UnknownTokens, end-to-end time, rework, quality
No matched comparison exists across coordinated and uncoordinated runs.
Historical sessions differ in task, model, and coordination maturity, so the current data cannot support a causal savings claim.
Read together
Every guarantee ships with its limit
Pairs as they appear in the beats. Citations point into Rally source at commit bc97e13.
| Beat | Rally guarantees | Rally does not |
|---|---|---|
| 01 · Presence | Rally's stated invariant is zero data loss: the ledger is append-only and canonical, and the SQLite cache is disposable and rebuilt from it. Append-only means never rewritten; the durability of the last line is the filesystem's promise, not Rally's.NORTH_STAR.md:29-31; docs/RALLY_ARCHITECTURE.md:70-75; RALLY.md:320,330; .gitignore:52-84 (ledger not distributed) | The name on a fact is self-asserted and unsigned. Presence proves a line was written, not who wrote it.RALLY.md:264-266; docs/security/TRUST-MODEL.md:86,113-117 |
| 02 · Reading, not receiving | Rally records and advises. It never gates, grants, schedules, spawns, retries, or executes. rally next returns a recommendation, not an execution order.NORTH_STAR.md:23-26; docs/TURN-LOOP.md:3,25,69; README.md:272 | Nothing is delivered to a peer on this path. A fact nobody reads changes nothing, and the word broadcast appears nowhere in Rally's docs. Injection into a managed session is a separate mechanism, shown in the systems view, and even there acceptance needs a target-authored fact.docs/COMMAND-SEMANTICS.md:28-29; docs/security/TRUST-MODEL.md:172-177; docs/AUTO-COORDINATION-HOOKS.md:268 |
| 03 · Claim | One owner per path, as coordination metadata, with a size-scaled lease: 30 minutes for a single file, 120 minutes for coarser or multi-scope claims.NORTH_STAR.md:31; crates/rally-cli/src/store.rs:5391-5411; crates/rally-cli/src/decay.rs:39-47,90-113; crates/rally-cli/src/lib.rs:2493-2545; docs/PROTOCOL-NORTH-STAR.md:345-353 (scope kinds) | A claim is not an operating-system permission boundary and Rally is not a sandbox. Nothing in the filesystem changed.docs/TURN-LOOP.md:67; docs/security/TRUST-MODEL.md:22-23 |
| 04 · Refusal | Overlapping claims are hard-refused at append time: an error, no fact appended, non-zero exit.crates/rally-cli/src/store.rs:5391-5411 | Rally refused the coordination record, not the file write. Nothing stops an agent editing src/api.ts anyway, and the before-write check is advisory by default.docs/TURN-LOOP.md:66-69; crates/rally-cli/src/check.rs:53-91; RALLY.md:37-40 |
| 05 · Reroute | Host-neutral: one CLI and one protocol across Claude, Codex, Gemini, and Cursor.NORTH_STAR.md:35; RALLY.md:65-76 | Rally does not judge whether the chosen lane is the right work. Its own docs say it cannot tell a useful task from a hostile one.docs/PROTOCOL-NORTH-STAR.md:30-32; dynamic-workflows/PROTOCOL.md:120-134 |
| 06 · Liveness | The durable log keeps the original claim fact forever; the live projection drops a claim only when a later release, resolve, session close, or reap closes it.crates/rally-cli/src/claim_authority.rs:71-84; RALLY.md:315-331 | Two Rally lookups can answer differently for the same path at the same moment. say claim refused; check before-write allowed with a warning. Idle over 900 s is advisory; destructive takeover needs 7200 s of silence.crates/rally-cli/src/store.rs:1732,1740; crates/rally-cli/src/check.rs:120-146; session observation 2026-09-12 |
| 07 · Retract | The log is never rewritten. The retraction itself stays visible and auditable, and room totals report retractions instead of hiding them.RALLY.md:242-270,348-419 | Rally does not judge which version is correct and does not chase the peer. A reader who read before the retraction and never reads again keeps acting on the withdrawn note.docs/PROTOCOL-NORTH-STAR.md:30-32; docs/COMMAND-SEMANTICS.md:28-29 |
| 08 · Handoff and receipt | A sender cannot close its own handoff, and a system or reaper action never closes it. Since commit 9129d81, retracted responses and system-authored closers do not count as proof of receipt.crates/rally-cli/src/store.rs:7413-7470,7452-7459,1147-1149; crates/rally-cli/src/lib.rs:12352-12363 | A receipt proves receipt, not completion, and the receiver's name on it is self-asserted and unsigned: Rally matches the target label and session, it does not authenticate anyone. A pane echo showing the prompt landed does not prove an LLM accepted it. The 30-day sweep is cleanup, not an answer; the item stays visible as stale in rally inbox.crates/rally-cli/src/store.rs:7431-7433; RALLY.md:264-266; docs/OPTIMIZATION-EVIDENCE.md:1-5,69-70; docs/COMMAND-SEMANTICS.md:110; crates/rally-cli/src/reaper.rs:35 |
| 09 · The boundary | Rally records the coordination facts but does not execute the task. Rally coordinates across agents; the verify loop drives within each lane.README.md:272; RALLY.md:130-138; dynamic-workflows/COORDINATION.md:118-138 | Rally does not know another tool's run state beyond what that tool self-reports. The observer admits its own limits too: heartbeat unknown, process liveness unknown.docs/OPTIMIZATION-EVIDENCE.md:1-5,69-70; run_status.py output 2026-09-12T07:46Z |
Accuracy matrix: every visible interaction mapped to a command, source, or test
Model tests assert the page's own scenario model, not Rally. Session observations are real command output captured 2026-09-12.
| Interaction | Command | Source at bc97e13 | Model test or session observation |
|---|---|---|---|
| Agent announces presence | rally enter --tool <tool>; rally say presence | crates/rally-cli/src/store.rs:527,601,631; docs/AGENT-STATE-MODEL.md:36 | model test: presence is a write |
| Agent reads the room | rally room --json; rally whoami; rally next | docs/COMMAND-SEMANTICS.md:28-32; docs/TURN-LOOP.md:25 | model test: no push events |
| Host hook reads on the agent's behalf | SessionStart and PreToolUse hooks | docs/AUTO-COORDINATION-HOOKS.md:268; docs/security/TRUST-MODEL.md:172-177; rally hooks status (enabled by default) | session observation: rally hooks status enabled by default, 2026-09-12 |
| Claim a scope with a lease | rally say claim --path <p> | NORTH_STAR.md:31; crates/rally-cli/src/decay.rs:39-47,90-113 | session observation: claim fact_8bd9_18d48373a53099e8 carried lease_expires_at |
| Overlapping claim refused at append | rally say claim (overlap) | crates/rally-cli/src/store.rs:5391-5411 | session observation: verbatim 2026-09-12: claim conflict ... overlaps the scope you requested |
| Advisory before-write finding | rally check before-write --path <p> [--strict] | crates/rally-cli/src/check.rs:53-91,120-146; RALLY.md:37-40 | session observation: 2026-09-12: allow:true with stale-owner-claim |
| Expired lease is not closure | rally doctor --reap-stale --apply | crates/rally-cli/src/claim_authority.rs:71-84; crates/rally-cli/src/reaper.rs:406-439 | session observation: 2026-09-12: three reap passes closed the expired claims |
| Retract without rewriting | rally retract <fact-id> --reason <why> | RALLY.md:242-270 | model test: withdrawal keeps the original |
| Handoff stays open until the target answers | rally say handoff --target <tool>; rally say receipt|resolve|artifact --ref <id> | crates/rally-cli/src/store.rs:7413-7470; crates/rally-cli/src/lib.rs:12352-12363 | session observation: receipt fact_fc26_18d483db677dce00 authored by the target |
| Receipt is not completion | rally say receipt | crates/rally-cli/src/store.rs:7431-7433 | model test: receipt and artifact are separate events |
| Build Loop run lifecycle | python3 scripts/run_status.py <run-id> | build-loop scripts/run_status.py; README.md:272 (Rally does not execute the task) | session observation: 2026-09-12T07:46Z: status abandoned, crash_signal stop_hook |
| Choose shared checkout or linked worktrees | rally run (default worktree) or rally run --shared / --no-worktree | crates/rally-cli/src/run_worktree.rs:4-34,81-131; RALLY.md:326-330 | model test: systems view layouts and fail-closed provisioning |
| Launch a managed lane | rally run claude|codex|opencode|gemini --backend auto --json; --dry-run first | RALLY.md:47-63,144-158; docs/HANDOFFS-AND-LAUNCHING-AGENTS.md:19-36 | model test: launch objects are separate and cited |
| Inject into a managed session | rally inject <session> --handoff <event-id> | RALLY.md:95-99,144-162; docs/HANDOFFS-AND-LAUNCHING-AGENTS.md:82-85 | model test: acceptance still needs target-authored evidence |
| Reach an unmanaged process | rally say handoff --target <tool>; rally watch | RALLY.md:95-99,160-162; docs/HANDOFFS-AND-LAUNCHING-AGENTS.md:86-110 | session observation: 2026-09-12: handoff 07:55:00Z, receipt 07:58:50Z by this unmanaged session |
| Daemon cutaway | rally daemon status|start|stop|serve | RALLY.md:320-330,420-452; RALLY.md:152-153; docs/HANDOFFS-AND-LAUNCHING-AGENTS.md:30,111 | session observation: 2026-09-12: daemon live false, commands still worked (fail open) |