Skip to content

Close the codex-limit edge cases left from item 47 (item 48) - #170

Merged
jposluns merged 5 commits into
mainfrom
item-48-flow-hardening
Oct 29, 2026
Merged

Close the codex-limit edge cases left from item 47 (item 48)#170
jposluns merged 5 commits into
mainfrom
item-48-flow-hardening

Conversation

@jposluns

@jposluns jposluns commented Oct 29, 2026 •
edited
Loading

Copy link
Copy Markdown
Owner

What changes

.claude/commands/flow.md closes the codex-limit edge cases left from item 47 (PR #169):

  • Failed codex legs. ORCH_MAX_FAILOVERS=1 means one account and no failover. A codex leg counts as a failed leg, not an unreachable family, whatever orch-verify advises, if it:
    • ends at that bound in any form (FAILOVER_LIMIT (exit 4), or an empty, degenerate, or 5xx chain);
    • ends EXHAUSTED (exit 5);
    • or is stopped by a cost cap.
  • Re-issue. A failed leg is re-issued at most once, after a wait and never with a higher limit. The re-issue happens at the first later FEED THE PIPELINE step at which orch-verify --pick codex succeeds and codex is not at its cost cap.
  • Owed seat. Until the re-issue delivers, the seat is owed, recorded as a ?HOLD on that QA step. Meanwhile the run does only pipeline top-up and read-only prep, and starts no other apply while the branch is open.
  • Held steps. A held SEED or COMBINE step is retried only when orch-verify --pick succeeds and the family is not at its cost cap.
  • Lost seats.
    • A genuinely unreachable family's seat goes first to a reachable, unbarred family with no seat. A same-family pass is used only where there is none.
    • Each lost seat gets its own substitute, named by the seat it fills.
    • When codex is barred or genuinely unreachable, its weighting passes to gemini.
  • Other rules.
    • A review that is a bypass hunt in whole or in part stays off codex.
    • "Fleet-rule substitution" and "item QA record" are defined where they are first used.
    • Smaller wording fixes, including a rewrap of the rounds sentence with its wording unchanged.

Not changed

The QA-round structure is unchanged, pending the separate DISCOVERY/VERIFY decision (item 46).

Known residuals, tracked as a hardening item

These LOW findings were deferred, not fixed:

  • If a seat's single re-issue also fails, the seat stays owed. This is deliberate, per the fleet rule; the step then falls to Blocked-verification.
  • The cycle step 9 reference applies only while work is in flight.
  • The length of the wait before the re-issue is undefined.
  • It is not stated whether an owed leg reviews the original revision or the fixed one.
  • A cost-capped gemini substitute has no handling.
  • A codex failure outside the listed endings has no handling.
  • The dual-family case where codex is barred and gemini is unreachable is not covered.

Verification

  • Seeds came from claude and gemini, combined by Fable. Every statement about orch-verify was checked against the installed /usr/local/bin/orch-verify.
  • Tri-family adversarial QA ran for 5 rounds with 4 fixing rounds. Round 5 on 7c2dca9:
    • codex PASS, with no findings;
    • gemini PASS, with no findings, and every quote and cited line verified at source;
    • claude FAIL, with LOW findings only, all deferred above.
  • All 24 CI gate commands exit 0 on 7c2dca9. The required checks pass on 7c2dca9, confirmed on two readings.

@jposluns
Close the codex-limit edge cases left from item 47 (item 48)
7edf0e9 
- A codex leg that exhausts its single failover without a result counts as
  unreachable. A leg stopped by a cost cap does not, and no leg is retried with
  a higher limit.
- A review that is a bypass hunt in whole or in part stays off codex.
- The terms "fleet-rule substitution" and "item QA record" are defined where
  they are first used.
- A held SEED or COMBINE step is recorded as a ?HOLD on that step and retried
  at each later FEED THE PIPELINE step.
- An unreachable family's seat goes first to a reachable, unbarred family with
  no seat. A same-family pass is used only where there is none, and each lost
  seat gets its own substitute.
- When codex is barred or unreachable, its weighting passes to gemini, or to
  the pass in gemini's seat if gemini is also lost.
- Smaller wording fixes: "the other seed family", "in the same way that", the
  placement of the third-family condition, and a rewrap of the rounds sentence
  (wording unchanged).
@jposluns
Correct the codex failover rule in /flow (item 48, round 1)
d69a5dd 
ORCH_MAX_FAILOVERS bounds the number of accounts orch-verify tries, so 1 means
one account and no failover. Reaching the bound is reported as an infra bound
(FAILOVER_LIMIT, exit 4), not exhaustion. The previous text wrongly counted it
as unreachable, which would have let codex be dropped over a quota.

- A codex leg that ends at that bound, or that a cost cap stops, is a failed
  leg, not an unreachable family. It is recorded and re-issued at most once,
  after a pause and never with a higher limit. If it is still undelivered,
  its seat is owed and is never dropped or substituted on that basis.
- A held SEED or COMBINE step is retried only when a seed family or a lawful
  combiner can run.
- The item QA record definition no longer carries an extra recording clause.
- In the case of two lost seats, the fallback pass holds the gemini seat and
  the fleet-rule substitution pass holds the codex seat.
@cloudflare-workers-and-pages

cloudflare-workers-and-pages Bot commented Oct 29, 2026 •
edited
Loading

Copy link
Copy Markdown

Deploying cleanlanguageai with  Cloudflare Pages  Cloudflare Pages

Latest commit: 7c2dca9
Status: ✅  Deploy successful!
Preview URL: https://be4ed779.ai-language.pages.dev
Branch Preview URL: https://item-48-flow-hardening.ai-language.pages.dev

View logs

@jposluns
Cover every codex failure ending in /flow (item 48, round 2)
2233d47 
- A codex leg that ends at its bound in any form (FAILOVER_LIMIT, or an empty,
  degenerate, or 5xx chain), ends EXHAUSTED, or is stopped by a cost cap is a
  failed leg, not an unreachable family, whatever orch-verify advises.
- Its single re-issue waits until `orch-verify --pick codex` succeeds. Until
  then, its seat is owed: a ?HOLD on that QA step, and the run moves to the
  next independent item.
- A held SEED or COMBINE step retries when a successful `orch-verify --pick`
  shows the family can run.
- The weighting passes to the gemini-seat fallback only if gemini is
  genuinely unreachable, and each fallback pass is named by the seat it fills.
@jposluns
Bound the owed codex seat and its retries in /flow (item 48, round 3)
2fe5fbf 
- While a codex seat is owed, the run does only pipeline top-up and read-only
  prep on other items, as cycle step 9 allows. It starts no other apply while
  this branch is open.
- Seed, combine, and codex retries need both a successful orch-verify --pick
  and the family not being at its cost cap. The codex re-issue is checked at
  each later FEED THE PIPELINE step.
- Codex is called "genuinely unreachable" in the tier text, matching the
  verifier-diversity paragraph.
@jposluns
Clarify the codex re-issue and seed wording in /flow (item 48, round 4)
7c2dca9 
- The single codex re-issue happens at the first later FEED THE PIPELINE step
  at which its condition holds, after a wait between attempts (not the PAUSE
  step), so it cannot be read as one re-issue per cycle.
- The second seed comes from "the seed family that can run".
- The two-lost-seats example says gemini is genuinely unreachable.
- The exit code is attached to FAILOVER_LIMIT, so the list of bound endings
  reads as two forms.
@jposluns
jposluns merged commit cfa0f2d into main Oct 29, 2026
5 checks passed
@jposluns
jposluns deleted the item-48-flow-hardening branch October 29, 2026 08:28
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@jposluns