DOTODO Workflow - Codex Project Manager

Workflow Token Savings

What changed when you fire off a DOTODO task today: Codex starts from the central task record, loads context progressively, stores bulky evidence as artifacts, and hands landing work to a fresh session.

1-3related tasks per run
900sreusable precheck cache
15-40%targeted savings
0safety rules removed
01
1Claim
LIVE
Start From
One Task
The old pattern often rediscovered the project. The current pattern claims a central task first, then lets that task decide what context is worth loading.
What happens first
1
Linear is the work source
task_pool_ops.py claim-next selects the next actionable issue, moves it to In Progress, and attaches the run id instead of scanning old TODO files.
2
The claim carries context
The task description, current run thread, recent notes, blocker history, and ready-transition reason become the first context window.
3
Only routed instructions load
Codex reads AGENTS.md and only the topic files that apply to the claimed task, instead of preloading every project rule.
less setup context before code work starts
readybefore backlog
1 PRcoherent scope
No guessingambiguous tasks are blocked or decomposed
02
2Load
Hydrate
Progressively
The workflow spends tokens only when evidence is needed: snapshots, narrow file reads, bounded agents, and compact command artifacts replace repeated broad exploration.
Modern context filters
1
Reuse project snapshots automation_snapshot.py captures repo, PR, task, and deploy state so follow-up runs do not re-read everything.
REUSE
2
Read only likely files Task provenance, named files, nearby tests, manifests, and routed docs come before broad source scans.
NARROW
3
Store bulky evidence outside chat evidence_artifact.py saves long command output, logs, diffs, and file slices with compact previews.
ARTIFACT
4
Poll with a helper bounded_poll.py replaces repeated manual status checks with one compact pass/fail summary.
BOUND
Today
Artifacts carry coordinates; the chat carries decisions.
Old way
Each run rediscovered state through repeated shell output.
+
Subagents stay bounded. They return findings and evidence references, not huge transcripts.
+
Palace stays useful. Recall is progressive, not artificially disabled to save tokens.
03
3Handoff
Stop At
The Boundary
Implementation, review, landing, and deploy no longer live in one ever-growing thread unless the task truly needs it.
End-of-run shape
A
Verification evidence is summarized
Tests, screenshots, conflict checks, PR status, and caveats are recorded without pasting full logs.
B
Follow-ups go to the task pool
New work is created through task_pool_ops.py, not buried in chat or repo markdown.
C
Landing starts fresh
The PR handoff carries task ids, PR URL, default action, deploy rationale, checks, and artifact paths for a new session.
the next run rehydrates from coordinates
What still never gets optimized away
Keep !
Safety and production boundaries
Destructive actions, production data, secrets, and deploy uncertainty keep their explicit checks.
Keep ?
Verification requirements
SpecTest, UI screenshots, high-risk review, and authenticated checks remain when the task calls for them.
Save -
Repeated rediscovery
Old transcripts, broad logs, polling loops, and duplicated PR context become compact references.