Explore Obelisk as a bounded historical-evidence provider for Replan #3792

Problem

When LoopX replans after repeated failures, exhausted hypotheses, or a handoff,
the current control-plane packet can say why replanning is required but may
not contain enough historical evidence to answer:

• did this agent already try the same route in an earlier session;
• which files, commands, hypotheses, or decisions actually failed;
• whether the proposed successor is materially new; and
• whether a blocker is new or merely rediscovered.

Obelisk indexes local Codex, Claude Code,
Kimi, Pi, and DSH session history into SQLite/FTS5 and exposes structured JS and
read-only SQL queries. It is a plausible source for this missing evidence, but
it must not become Replan authority or a raw-transcript injection path.

Evidence and product judgment

Obelisk's own public evaluations define a useful but narrow boundary:

• #39 reports strong
  conversational and temporal retrieval, and a reduction in repeated dead-end
  commands from 6.7% to 0.5%, but no statistically significant SWE retry uplift
  from raw trajectory retrieval. Retrieved context was also large (25–51k
  characters per query).
• #46 reports no measurable
  benefit from foreign-agent cross-task archives, while same-agent,
  same-repository longitudinal history showed promising efficiency and quality
  signals.

This supports a cold-path historical-evidence provider, not retrieval on
every Turn and not an automatic memory writer.

Proposed pilot

Add a default-off optional provider, tentatively
obelisk-session-evidence, under the existing agent-turn-recall capability.
Call it only for typed situations such as:

1. repeated or materially unchanged Replan attempts;
2. exhausted hypotheses with missing historical coverage;
3. handoff or resume where earlier session evidence is unavailable; or
4. an explicit operator request for prior-session evidence.

The first slice should be read-only and limited by default to the same agent,
same project/repository, recent or revision-relevant sessions. It should return
a bounded evidence packet plus a content-free receipt containing provider and
index revision, scope, provenance, opaque session/message ids, evidence type,
freshness/truncation, omitted count, and query/result digests.

Retrieval never satisfies Replan by itself. Existing typed outcomes still have
to establish a new surface, hypothesis, probe, grounded successor, blocker, or
coverage-backed exhaustion. Every result and receipt must declare
grants_new_action_authority=false and external_writes_performed=false.

Placement and licensing

• Capability owner: existing agent-turn-recall; no new generic Obelisk
  capability.
• Provider id: obelisk-session-evidence.
• Delivery: optional extension/package, disabled by default.
• Boundary: invoke an independently installed, unmodified Obelisk CLI through
  a bounded structured adapter. Do not vendor or link AGPL-3.0-only Obelisk
  code into Apache-2.0 LoopX core.
• Prefer an upstream structured JSON query envelope over arbitrary generated
  JavaScript. A fixed escaped query template is the maximum acceptable MVP.

Pilot acceptance criteria

• Lower repeated-dead-end rate on matched, same-repository Replan cases.
• Lower tokens/provider calls or time-to-grounded-successor without reducing
  task success or evidence quality.
• Bounded context size, deterministic scope enforcement, explicit truncation,
  and fail-closed rejection of unsupported query options.
• No raw transcripts, local paths, identities, credentials, or query text in
  public LoopX projections.
• Index freshness and revision are observable; stale or incomplete indexes are
  explicit, not silently treated as complete history.
• Cross-agent retrieval and memory writes remain out of scope unless separately
  enabled and qualified.

No-go criteria

Do not promote the provider if matched evaluation shows no measurable reduction
in repeated work or efficiency, if retrieval harms success/evidence quality, if
results cannot be bounded reliably, or if scope/query failures can broaden the
search silently.

Known upstream risks to qualify

• CI coverage: obelisk#34
• unsupported options may fail open: obelisk#94
• indexing/finalization cost: obelisk#75, obelisk#105
• tokenizer false positives: obelisk#76
• stale same-mtime Codex updates: obelisk#104

The accompanying RFC should define the typed request/result/receipt contract,
activation and disable lifecycle, staged evaluation, privacy boundary, and
relationship to Replan, Explore, Reward Memory, and OpenViking.