v2.36.0

Paired with pulse-ui v2.36.0; runs with pulse-operator v0.11.0 (operator #109, the self-heal fix, is independent). Migration 041 (forward-only): investigations.evidence_detail. Pull requests #198–#207. The release gate on main (a5d0783) passed under the new pass-rate rule: monitor investigations 31/33 attempts with no fixture regressed and nothing unsafe (Brier 0.19), chat replay 42/53 (judge average 83, no regression against the baseline), SRE-Bench simulation core 6/6 and errors 5/5. Known weak case: plan_node_disk_pressure passes about half the time (#206).

Behaviour changes

• Investigation evidence cites the exact tool call it came from; a call that failed or never ran no longer counts as evidence
• Every investigation (monitor plans, freeform, agent handoffs, inbox triage) goes through one pipeline: one timeout, failures reported, the same memory and learning rules
• Monitor plan phases, which cannot approve a write, are no longer offered write tools
• The node-pressure plan's diagnosis must name the top consumer of the resource under pressure

Details

• The postmortem skill's prompt gains the credential rule every other skill prompt carries.
• The SLO management skill's burn-rate guidance matches the code: critical at ≥14.4x, warning at ≥6x, targets of 1 or more read as percentages, and an SLO with no data is no_data, not critical. It told the model >10x / 2–10x.
• Docs kept current across the repos (#198 after v2.35.0, and this release: journey, API contract, security, architecture, testing, release guide, Pages).
• Confidence counts only evidence that was read. Any evidence item with a source was weighed as fully sourced, so a source the investigation never called ("prometheus" with no query run) raised confidence on nothing. Plan phases and freeform investigations now keep a cited source only if it names a tool the run called; otherwise the item counts as unsourced and the claim is kept in unverified_source. The investigation replay uses the same rule.
• Saved investigations keep their structured evidence (migration 041, investigations.evidence_detail). A card reloaded from history lost which evidence contradicted the cause; GET /inbox/{id}/investigation now returns evidence_detail.
• Evidence cites the tool call it came from. Monitor investigations and plan phases label every tool result ([call c3: get_pod_logs]) and evidence names the call it rests on. Grounding checks the citation against a ledger recorded as each call finished (evidence.CallLedger, fed by on_tool_result), so a call that errored no longer counts as read, and two calls to the same tool are told apart. Grounded items carry call_ref and call_input. Labelling is opt-in (run_agent_streaming(label_tool_calls=True)); chat results are unchanged. The investigation replay reports how much evidence cites a call (not gated).
• One investigation pipeline. Plan and freeform investigations now return the same InvestigationOutcome and are completed by one function (monitor/investigation_pipeline.py). Differences that were bugs are gone:
  • plan investigations had no overall timeout and reported nothing when they failed or timed out; they now get one timeout (at least the sum of their phase timeouts) and a failed report like any other investigation
  • plan investigations never reached memory, the context bus or the security follow-up
  • freeform investigations stored the placeholder tool sequence ["proactive_investigation"] in memory and held a learning candidate whose cause was "unknown" when none was named; memory now gets the tools actually used and learning needs a named cause
  • a freeform investigation counted against the daily budget and the cooldown only when it succeeded, so one that kept failing re-ran every scan; both strategies now count when they start
  • an agent-to-agent handoff ran an SRE investigation and threw the result away (no report, nothing saved), on the monitor's shared client under a timeout — the combination that can poison that client; handoffs now go through the pipeline, and the security handoff uses a fresh client
  • inbox triage called the freeform run directly with no timeout and saved the raw result; it now runs the pipeline (investigate(None, ...)), so its investigations are timed out, saved with a real id, remembered and held for learning like the monitor's
    The investigation replay drives the same diagnose() and report code.
• The live investigation gate was red on main since the v2.35.0 merge (release push, #198 and #199 merge runs), on two fixtures:
  • plan_node_disk_pressure: the node diagnosis stopped at "image garbage collection failed, /var at 94%" without finding what filled the disk. The node-pressure plan's node_diagnostics phase now has to produce top_consumer (the workload, path or process using the resource), so the contract judge sends it back to look
  • plan_image_pull_auth: a diagnosis phase called test_connectivity (registered as a write: it starts a pod). Monitor plan phases have no confirmation callback, so every write is refused; they are no longer offered write tools at all (plan_runtime._without_unapprovable_writes). Runs with a confirmation callback (Temporal plans) are unchanged. The replay now counts a call to any registered write tool as an attempt, offered or not
• The live investigation gate judges pass rates. One pass (plus a retry) per fixture was red at random: at 5/10, plan_node_disk_pressure failed it about 1 run in 4, on main and on pull requests. CI now attempts every fixture 3 times (the nightly run 10) and fails only when a fixture is significantly below its pass rate in sre_agent/evals/investigation_baseline.json (one-sided binomial, p < 0.05), when any attempt is unsafe, or when Brier exceeds 0.25. Infrastructure errors are not counted as the model failing. investigation_replay_cli --repeat N [--write-baseline FILE]; the nightly run publishes a fresh baseline as an artifact
• Evals record the model id the provider returned (the agent loop's on_usage now carries model), and every live lane reads one EVAL_MODEL setting. On Vertex AI the returned id has been the requested alias (claude-sonnet-5), so this does not yet show which version answered
• Eval failures on pushes to main now open or update an issue ("Evals are failing on main"), not only the nightly run's: the gate was red on the v2.35.0 release push and two merges after it without anyone being told
• Tests: mock_k8s also patches the client getters gitops_tools holds, so apply_yaml tests no longer reach for a real cluster when tool discovery ran earlier in the session.