orchestrator

adam/orchestrator

Archived

Fork 0

Commit graph

Author	SHA1	Message	Date
Adam Moussa	e7ef4387e6	fix(pipeline): single-shot Claude calls + planner actually reads review findings Two root causes behind 'every task parks, and slowly': 1. SINGLE-SHOT INVOKER: subscription Claude calls ran as 40-turn, tool-enabled agentic sessions (--max-turns 40, $2 budget) for what are pure reasoning->JSON completions — minutes-long, and Claude wandered/returned unparseable output. _DEFAULT_MAX_TURNS 40->1 + allowed_tools=[] -> fast deterministic single turn. 2. PLANNER FEEDBACK KEY MISMATCH: the review stage writes verdict/findings, but _format_review_feedback read decision/notes/comment (never present) -> the planner re-planned with EMPTY feedback, re-introduced the rejected flaw ('assumptions persist'), hit the review cap, parked. Now reads verdict/findings (old keys kept as fallback) so GPT-4.1's objections reach the re-plan. Plus: park notifications infer the phase the task was IN (review/plan/clarify) instead of the terminal 'parked'. Tests: planner real-verdict-keys regression + coordinator phase-inference. 1149 pass.	2026-06-23 15:49:40 -04:00
Adam Moussa	15a416d31a	Add Plane-2 leaf scaffold (pipeline graph, nodes, HITL, transports, CI) Consolidates the 18 leaf modules from the r720-plane2-scaffold workflow onto the foundation commit. Full suite: 535 passed, 1 skipped; ruff + format clean. Built (pre-deployment scaffold only — nothing provisioned/enabled): - LangGraph pipeline graph.py (INTAKE->CLARIFY->PLAN, interrupt()/resume, checkpointer-injectable) - nodes: clarifier (98% gate), planner, review_loop (GPT-4.1), builders->candidate diff, verifier - §3.3.1 HITL: ledger ops, resume_worker, deadline_timer, recovery sweep, responder - transports: slack / github / claude_code adapters - ci_gate (pure-code pass/fail), operator_cli, run-team.py entry, P1 sim harness - ci/agent-team-apply-verify.yml (split untrusted/privileged jobs) — authored, disabled KNOWN OPEN FINDINGS (verifier/cross-review, not yet fixed — see follow-up): - builders denylist: 4 execution-proven bypasses (delete, mode-change, copy-to, out-of-scope delete) - §3.3.1 CAS: BEGIN IMMEDIATE outside try/except; shared-connection txn nesting unsafe under concurrency - operator_cli: missing re-deliver/force-resume; audit-after-mutate ordering gap - ci yaml: GPT-4.1 cross-review PASS w/ 4 FIX items (symlink path escape, etc.) - P1 sim harness models the ledger layer, not real LangGraph interrupt/resume; P1 exit criteria not yet truly proven Deploy-gated (NOT done): IAM/step-ca/Roles Anywhere/confluence-bot provisioning, /sh-security-review sign-off, live Slack/CI, rsync, live dry-runs, Adam approval.	2026-06-17 15:16:12 -04:00

Author

SHA1

Message

Date

Adam Moussa

e7ef4387e6

fix(pipeline): single-shot Claude calls + planner actually reads review findings

Two root causes behind 'every task parks, and slowly':
1. SINGLE-SHOT INVOKER: subscription Claude calls ran as 40-turn, tool-enabled
   agentic sessions (--max-turns 40, $2 budget) for what are pure reasoning->JSON
   completions — minutes-long, and Claude wandered/returned unparseable output.
   _DEFAULT_MAX_TURNS 40->1 + allowed_tools=[] -> fast deterministic single turn.
2. PLANNER FEEDBACK KEY MISMATCH: the review stage writes verdict/findings, but
   _format_review_feedback read decision/notes/comment (never present) -> the
   planner re-planned with EMPTY feedback, re-introduced the rejected flaw
   ('assumptions persist'), hit the review cap, parked. Now reads verdict/findings
   (old keys kept as fallback) so GPT-4.1's objections reach the re-plan.
Plus: park notifications infer the phase the task was IN (review/plan/clarify)
instead of the terminal 'parked'.
Tests: planner real-verdict-keys regression + coordinator phase-inference. 1149 pass.

2026-06-23 15:49:40 -04:00

Adam Moussa

15a416d31a

Add Plane-2 leaf scaffold (pipeline graph, nodes, HITL, transports, CI)

Consolidates the 18 leaf modules from the r720-plane2-scaffold workflow onto
the foundation commit. Full suite: 535 passed, 1 skipped; ruff + format clean.

Built (pre-deployment scaffold only — nothing provisioned/enabled):
- LangGraph pipeline graph.py (INTAKE->CLARIFY->PLAN, interrupt()/resume, checkpointer-injectable)
- nodes: clarifier (98% gate), planner, review_loop (GPT-4.1), builders->candidate diff, verifier
- §3.3.1 HITL: ledger ops, resume_worker, deadline_timer, recovery sweep, responder
- transports: slack / github / claude_code adapters
- ci_gate (pure-code pass/fail), operator_cli, run-team.py entry, P1 sim harness
- ci/agent-team-apply-verify.yml (split untrusted/privileged jobs) — authored, disabled

KNOWN OPEN FINDINGS (verifier/cross-review, not yet fixed — see follow-up):
- builders denylist: 4 execution-proven bypasses (delete, mode-change, copy-to, out-of-scope delete)
- §3.3.1 CAS: BEGIN IMMEDIATE outside try/except; shared-connection txn nesting unsafe under concurrency
- operator_cli: missing re-deliver/force-resume; audit-after-mutate ordering gap
- ci yaml: GPT-4.1 cross-review PASS w/ 4 FIX items (symlink path escape, etc.)
- P1 sim harness models the ledger layer, not real LangGraph interrupt/resume; P1 exit criteria not yet truly proven

Deploy-gated (NOT done): IAM/step-ca/Roles Anywhere/confluence-bot provisioning,
/sh-security-review sign-off, live Slack/CI, rsync, live dry-runs, Adam approval.

2026-06-17 15:16:12 -04:00

2 commits