| name | atlas-abstract |
| description | Phase A (Abstract) — AgentFold compression of the Locate trajectory into a ≤2000-token working-memory summary plus a Memex excerpt store. Runs in a clean-context subagent to prevent trajectory contamination. Mechanically validates that every FINDING-XXX, GAP-XXX, and ESCALATION_TRIGGER is preserved. Use after findings.md is complete and before synthesis. Triggered automatically at phase boundary, or at ≥60% context-used telemetry threshold. |
| allowed-tools | memex_read view_file |
| metadata | {"methodology":"ATLAS","phase":"A"} |
SKILL: Abstract — AgentFold & Memex (Phase A)
When to use
Load when: findings.md is complete and you are about to Synthesize.
Unload when: the fold summary has passed mechanical validation and Phase S
has begun.
Contract
| Field | Value |
|---|
| LLM calls permitted | Yes — this phase is summarization. But see the clean-context rule below. |
| Tool budget | ≤ 5% of mission max_tool_calls |
| Output | (1) a ≤2000-token fold summary in working memory; (2) a persisted Memex index |
Why this phase exists
Long-horizon agents fail predictably when the Locate trajectory balloons
beyond the attention budget. Step-wise summarization loses critical
constraints; fixed truncation destroys intermediate reasoning; no summarization
at all courts context rot.
AgentFold is the alternative: treat Locate as a branch that you
explicitly return from. The return folds the branch into a self-authored
summary whose contract is mechanically checked. The Memex holds the raw
evidence, so the fold can be lossy without being destructive.
The clean-context rule
The fold is performed by a prompt that has seen:
mission.mdmap.mdfindings.md
And nothing else. Specifically, no Locate transcript, no tool logs, no
intermediate reasoning. Why: trajectory contamination. If the fold prompt
remembers the path it took, it will preserve path-relevant details at the
expense of mission-relevant details. The fold is for the mission, not the
journey.
Implementation: spawn an ephemeral subagent whose only inputs are the three
artifacts above and whose only output is the fold summary. Your current
(parent) context keeps the returned summary; discard its working state.
Fold contract
The summary MUST preserve, in this order:
1. Mission header
MISSION-ID: <id>
DECISION_TARGET: <verbatim from mission.md>
SCOPE: <verbatim globs>
STATUS: completed | partial | blocked
2. Decision-ready answer
A ≤200-token prose answer to DECISION_TARGET, with inline FINDING-XXX
references after each factual clause. Example:
RecordVote#call is the sole writer to cast_vote_records [FINDING-017].
It delegates authorization to VotingAuthorizer#authorize_ballot
[FINDING-017]. No Sidekiq path exists [FINDING-021, GAP-003].
3. Findings index
A flat list. No claim text repeated — just IDs, tiers, and anchors.
FINDING-017 | H | app/flows/vote_casting/record_vote.rb:42-78
FINDING-017 | H | app/policies/voting_authorizer.rb:12-40
FINDING-021 | M | app/flows/vote_casting/** (ruled_out)
...
4. Open questions / gaps
Every GAP-XXX from findings, verbatim ID and one-line status.
5. Escalation log
Every ESCALATION_TRIGGER that fired during Locate, with timestamp and
resolution. (Especially important for election-critical or
security-sensitive missions where triggers map to hard gates.)
What MAY be dropped
- Intermediate tool outputs (they're in Memex if needed).
- Failed probes that produced no finding (the ones that led to
ruled_out notes are preserved via their finding).
- Narrative reasoning ("I first thought X, but then realized Y"). Downstream
agents want conclusions, not process.
- Duplicate anchors across findings — listed once in the index, referenced
from the answer prose.
Mechanical validation
Before accepting the fold, run these checks (regex-grade, deterministic):
assert every FINDING-XXX from findings.md appears in the fold index.
assert every GAP-XXX from findings.md appears in §4.
assert every ESCALATION_TRIGGER from mission.md §ESCALATION_TRIGGERS
that fired during Locate appears in §5.
assert §2 (decision-ready answer) ≤ 200 tokens.
assert total fold ≤ 2000 tokens.
assert every sentence in §2 is followed by at least one FINDING-XXX reference.
If any check fails, the mission was under-scoped — do not retry the fold
with a different prompt. Halt and escalate to human or spawn a follow-up
mission.
Memex hygiene
What goes in
- Every raw excerpt referenced by a FINDING.
- Key intermediate queries (e.g., the symbol-probe result graph for
CastVoteRecord) — byte-exact, not summarized.
- The raw
findings.md (for audit).
Keying
Content-addressable, SHA-256 over the raw bytes. Example:
memex://excerpt/7f31ab...
Retrieval contract
memex.read(ref) returns bytes identical to what was written. No
transformation, no summarization at read time. This is what lets downstream
agents re-inject exact syntax into their context if needed.
What does NOT go in
- LLM-authored summaries (those belong in working memory).
- Tool transcripts (unless you deliberately want audit trail — expensive).
- User PII or secrets scraped during exploration. Memex is per-mission;
assume any content could be surfaced to downstream agents.
Telemetry
Record the fold ratio:
fold_ratio = tokens(fold_summary) / tokens(locate_trajectory)
Target: ≤ 0.1 (≥90% compression). The AgentFold literature shows this is
achievable and that violating it usually means the Locate phase collected
too much. Feedback signal: if you repeatedly hit fold_ratio > 0.2 across
missions, revisit your Locate skill — you're probably reading too much.
Exit gate
