Add chain-stall escalation to Themis orchestrator #53

New Issue

2026-05-30T11:04:50+02:00

product-owner commented

2026-05-30 11:04:50 +02:00

WIG: WIG-3 (Complete agent orchestration and MCP integration)
Compound-RICE: reach=20 impact=4 confidence=0.95 synergy=2.0 effort=3 maintenance=1 -> score=50.7

Problem

The F1.2 dependency chain (#35->#50->#44->#49->#51) has had zero PR activity for at least 6 consecutive Themis cycles. Themis currently raises an "escalation flag" in its comment but has no formal protocol to force a human decision or decompose the blocking issue. The result is that agents keep iterating the same RICE ranking with no new information, burning cycle budget without progress.

Proposed Solution

In the Themis cycle script, after computing RICE rankings, count consecutive cycles where the top-N vital issues have had no PR created or closed (read from Gitea PR API, no new field needed)
When count >= 3, auto-create a labelled issue: type/stall-review, body = chain summary + blocking reasons + "human decision required: decompose, deprioritise, or assign directly"
Post a comment on the head blocking issue linking the new stall-review issue
Label the head issue status/stalled to signal agents to skip pickup
Gate: do not create a second stall-review if one is already open for the same chain

Acceptance Criteria (mini-UAT)

  Given the Themis orchestrator is run and the top-3 vital issues have no PR activity in the last 3 cycle records
  When Themis completes its ranking pass
  Then a new issue is created with label `type/stall-review` containing the chain summary, cycle count, and blocking reasons
  And a comment linking to the new issue is posted on the head blocking issue (#35)
  And the head issue receives label `status/stalled`
  And if a `type/stall-review` issue is already open, no duplicate is created
  And Themis cycle output includes "STALL_ESCALATED #NNN" in its summary line

Marketing hint: Turns a silent infinite loop into a human decision point -- agents stop spinning, humans get a clean escalation ticket.

Suggested agent: @adf:themis-orchestrator

Created by @adf:product-owner (Themis) -- Compound-RICE + Essentialism + UAT cycle 2026-05-30

**WIG**: WIG-3 (Complete agent orchestration and MCP integration) **Compound-RICE**: reach=20 impact=4 confidence=0.95 synergy=2.0 effort=3 maintenance=1 -> score=50.7 ## Problem The F1.2 dependency chain (#35->#50->#44->#49->#51) has had zero PR activity for at least 6 consecutive Themis cycles. Themis currently raises an "escalation flag" in its comment but has no formal protocol to force a human decision or decompose the blocking issue. The result is that agents keep iterating the same RICE ranking with no new information, burning cycle budget without progress. ## Proposed Solution - In the Themis cycle script, after computing RICE rankings, count consecutive cycles where the top-N vital issues have had no PR created or closed (read from Gitea PR API, no new field needed) - When count >= 3, auto-create a labelled issue: `type/stall-review`, body = chain summary + blocking reasons + "human decision required: decompose, deprioritise, or assign directly" - Post a comment on the head blocking issue linking the new stall-review issue - Label the head issue `status/stalled` to signal agents to skip pickup - Gate: do not create a second stall-review if one is already open for the same chain ## Acceptance Criteria (mini-UAT) ```gherkin Given the Themis orchestrator is run and the top-3 vital issues have no PR activity in the last 3 cycle records When Themis completes its ranking pass Then a new issue is created with label `type/stall-review` containing the chain summary, cycle count, and blocking reasons And a comment linking to the new issue is posted on the head blocking issue (#35) And the head issue receives label `status/stalled` And if a `type/stall-review` issue is already open, no duplicate is created And Themis cycle output includes "STALL_ESCALATED #NNN" in its summary line ``` **Marketing hint**: Turns a silent infinite loop into a human decision point -- agents stop spinning, humans get a clean escalation ticket. **Suggested agent**: @adf:themis-orchestrator --- *Created by @adf:product-owner (Themis) -- Compound-RICE + Essentialism + UAT cycle 2026-05-30*

product-owner referenced this issue

2026-05-30 11:05:25 +02:00

feat: wire ExitCode into main() return for f1.2 robot exit code contract #35

product-owner referenced this issue

2026-05-31 07:58:39 +02:00

feat: wire ExitCode into main() return for f1.2 robot exit code contract #35

product-owner referenced this issue

2026-06-01 02:58:44 +02:00

stall-review: F1.2 robot-contract chain stalled 8 cycles -- force decision (assign #35 / de-scope / wire execution) #55

product-owner commented

2026-06-01 02:59:23 +02:00

Themis cycle 2026-06-01 02:55Z -- #53 protocol manually executed (cycle 8)

This issue (#53) proposed: "when stall count >= 3, auto-create a stall-review issue forcing a human decision." The feature is not yet implemented, so I have executed the protocol manually: see #55 (stall-review).

Why now, and why not another RICE comment: the F1.2 chain (#35->#50->#44->#49->#47/#51) has been the vital few for 8 consecutive cycles. New, decisive evidence this cycle: agent-tasks has had zero PRs, ever -- the backlog has no execution arm. The bottleneck is confirmed to be execution-pickup, not prioritisation, so re-ranking confidence has collapsed (~0.1). Doing the 9th scoring ritual means not forcing the decision; I chose to force it.

This validates #53's own value: a passive escalation flag in a comment changed nothing across 6+ cycles. The protocol needs to (a) create a tracked, labelled artifact and (b) name a concrete decision with an owner above the product-owner mandate. #55 demonstrates the target behaviour for the eventual implementation.

Compound-RICE (unchanged, still critical): #53 = 50.7. Recommend #53 itself be picked up next so this escalation is automated rather than hand-run each cycle.

Refs #55 #54 #35

## Themis cycle 2026-06-01 02:55Z -- #53 protocol manually executed (cycle 8) This issue (#53) proposed: "when stall count >= 3, auto-create a stall-review issue forcing a human decision." The feature is not yet implemented, so I have **executed the protocol manually**: see **#55** (stall-review). **Why now, and why not another RICE comment:** the F1.2 chain (#35->#50->#44->#49->#47/#51) has been the vital few for 8 consecutive cycles. New, decisive evidence this cycle: `agent-tasks` has had **zero PRs, ever** -- the backlog has no execution arm. The bottleneck is confirmed to be **execution-pickup, not prioritisation**, so re-ranking confidence has collapsed (~0.1). Doing the 9th scoring ritual means not forcing the decision; I chose to force it. **This validates #53's own value**: a passive escalation flag in a comment changed nothing across 6+ cycles. The protocol needs to (a) create a tracked, labelled artifact and (b) name a concrete decision with an owner above the product-owner mandate. #55 demonstrates the target behaviour for the eventual implementation. **Compound-RICE (unchanged, still critical)**: #53 = 50.7. Recommend #53 itself be picked up next so this escalation is automated rather than hand-run each cycle. Refs #55 #54 #35

kimie05c34be198a20b9 referenced this issue

2026-06-26 21:18:33 +02:00

ADF Progress Check: 2026-06-26 #60

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: terraphim/agent-tasks#53