Context items reopened
36 1
Bounded reconstruction burden reduced
Component Architecture Evolution
From session-bound execution to durable workflow continuity
Proven mechanically; human value unresolved.
Optimize-Worker now preserves bounded workflow state across interruption, forced crash and a genuine Codex session boundary. The evidence proves mechanical continuity—not universal speed, human productivity or production reliability.
Implemented on canonical mainBounded stress completeNo overall operational pass
Executive Summary
The smallest sufficient response to the demonstrated problem
We built a bounded durable-checkpoint primitive before adopting LangGraph because the demonstrated problem was continuity loss after interruption—not graph topology or distributed orchestration.
The thin checkpoint solved the mechanical problem: reopened context fell from 36 items to 1, repeated completed steps fell from 5 to 0, observed state loss fell from 4 to 0, and the workflow recovered through a forced crash and a genuine cross-Codex-session resume.
LangGraph remains parked pending stronger complexity triggers and bounded net-benefit proof. This page documents mechanical proof, not a production recommendation. Human operational value remains unresolved because the 137-second observation represents fresh Codex re-entry, not a controlled human baseline.
- Runtime strategy
- Retain thin durable checkpoint
- LangGraph
- Parked, not rejected
- Reconsideration
- At least two material triggers plus bounded net-benefit proof
- Human value
- Unresolved
Architecture Evolution
The architecture changed at the point of failure: workflow state, not orchestration breadth.
Before
Session-bound execution
Interruption
Manual context reconstruction
Repeated-step and state-loss risk
After
Durable checkpoint runtime
Persisted SQLite WAL state
Crash and cross-session resume
Continue from last valid state
Repeated completed steps
5 0
No repeated completed step observed
Observed live-state loss
4 0
No state loss observed in the bounded run
Artifact completeness
Not established 71/71
Authenticated stress corpus recovered completely
From → To → Result
From
Architecture intention for central orchestration
To
Canonical bounded durable-checkpoint primitive
Result
Implementation exists on canonical main
From
Session-local workflow state
To
SQLite WAL-backed persisted checkpoint state
Result
Pause and resume from persisted state
From
Manual context reconstruction
To
Checkpoint-assisted restoration
Result
Context items reopened 36 → 1 in the paired experiment
From
Completed work vulnerable to repetition
To
Persisted completed-step tracking
Result
Repeated completed steps 5 → 0
From
Live state vulnerable to interruption
To
Versioned persisted workflow state
Result
Observed state loss 4 → 0
From
Normal interruption proof only
To
Worker stop, forced crash and expired-lock recovery
Result
No false completed state after the forced crash
From
Same-session continuity
To
Genuine cross-Codex-session resume
Result
Workflow reached COMPLETED, checkpoint version 6
From
Small evidence bundle
To
Authenticated 71-file PM Brain stress corpus
Result
Artifact completeness 71/71
From
Implicit side-effect behavior
To
Caller-controlled durable side-effect marker
Result
Side effect remained 1 → 1; runtime idempotency not proven
From
Possible LangGraph adoption
To
Thin custom runtime retained
Result
LangGraph parked pending evidence triggers
From
Human value implied by automation
To
Fresh Codex re-entry measured separately
Result
137 seconds; controlled human value remains unresolved
Evidence that changed the architecture
The paired run established mechanical continuity improvement.
The stress extension authenticated 71 artifacts across three interruption points.
A forced process crash preserved REVISION_REQUIRED rather than a false completed state.
A separate process recovered an expired lock and persisted WAITING_APPROVAL.
A genuinely new Codex session resumed and completed checkpoint version 6.
Stale/corrupt cases were detected; the caller-controlled marker prevented repetition.
Runtime Recovery Flow
◆ state checkpoint · ⚠ failure point · ✓ recovery
Worker
◆ checkpoint v3
Waiting review
review begins
Revision required
⚠ forced crash, exit 97
Waiting approval
✓ lock recovery, v5
Completed
✓ new Codex session, v6
Why not LangGraph yet
The current evidence shows a continuity problem that the smaller checkpoint primitive addresses. It does not yet demonstrate dynamic graph complexity, distributed execution, replay complexity or maintenance burden sufficient to justify framework migration. LangGraph is parked—not rejected—and remains eligible when evidence changes.
LangGraph Decision
New workflow need
Evaluate current runtime
Thin checkpoint sufficient?
Yes: retain
At least 2 material triggers?
No: retain · Yes: spike
Bounded spike shows net benefit?
Only then: separate adoption decision
Decision Criteria
CriterionCurrent evidenceDecision implication
Demonstrated problemState continuity after interruptionThin checkpoint directly addresses it
Dynamic branching, cycles or subgraphsNo material complexity demonstratedNo graph framework required yet
Distributed executionNot currently required; local locking onlyDo not pre-build distributed orchestration
Pause/resume and recoveryProven in bounded forced-crash and cross-session casesRetain current runtime
State-model maintenance burdenNo material burden demonstratedAdditional abstraction not justified
Human-in-the-loop complexityCurrent review/revision/approval loop remains manageableNo material trigger yet
Replay or time travelNo demonstrated operational requirementDefer framework expansion
External side-effect idempotencyCaller-controlledMust be solved explicitly regardless of framework
Human/operator valueNo controlled human baselineGather evidence; do not change framework to solve a measurement gap
Migration net benefitNo comparative evidenceLangGraph remains parked
Reconsideration Triggers
- 01
Dynamic branches, cycles or subgraphs create measurable defects or maintenance burden.
- 02
Multiple remote workers require distributed leases or concurrent ownership.
- 03
Approval, interruption, revision and escalation semantics exceed the current state model.
- 04
Replay, time travel or partial re-execution becomes an operational requirement.
- 05
Custom checkpoint maintenance or test cost exceeds the likely migration cost.
- 06
Existing telemetry cannot explain execution path, state transition or failure origin.
- 07
Workflow types and integrations create demonstrable orchestration sprawl.
- 08
A bounded LangGraph spike demonstrates measurable net benefit.
AT_LEAST_TWO_MATERIAL_TRIGGERS
+ BOUNDED_SPIKE_SHOWS_MEASURABLE_NET_BENEFIT
+ MIGRATION_RISK_IS_ACCEPTABLE
+ SEPARATE_ARCHITECTURE_AUTHORITY_IS_GRANTEDThis is a bounded governance rule for this decision, not an industry standard.
Proven / Not Proven
Proven within bounded evidence
- Persisted pause/resume works.
- Forced-crash recovery preserved a valid non-completed state.
- Genuine cross-Codex-session resume completed.
- Completed-step repetition and observed state loss were zero.
- Stale/corrupt checkpoint detection passed bounded cases.
- Artifact completeness was 71/71.
- A caller-controlled side-effect marker prevented repetition.
Not proven — an honest evidence boundary
- Controlled human productivity improvement or universal speed improvement.
- Business value or production reliability.
- Distributed safety.
- Runtime-guaranteed external-side-effect idempotency.
- Full Big Crew workflow readiness.
- Superiority over LangGraph.
Current Architecture Status
CANONICAL_BOUNDED_DURABLE_CHECKPOINT_IMPLEMENTEDMECHANICAL_CHECKPOINT_STRESS_CRITERIA_PASSEDHUMAN_OPERATIONAL_VALUE_UNRESOLVEDLANGGRAPH_PARKEDOptimize-Worker has a canonical bounded recovery capability. Its mechanical behavior survived the tested interruptions. The evidence does not yet justify a human-value claim, a production-reliability claim or a framework migration.
Evidence References
View public-safe evidence identities
- Canonical implementation — commit
e902d336…, treeacdfc7a9… - Paired real-workload experiment — 36→1, 5→0, 4→0
- Bounded stress validation — 71/71 artifacts
- Cross-session result — COMPLETED, checkpoint v6
- Final interpretation — mechanical criteria passed; human value unresolved; no overall pass
- KB custody — local commit preserved; remote inclusion not claimed