| name | deep-reflect |
| description | Deep research over the Egregore org memory — multi-hop, evidence-verified synthesis of what the org collectively knows — when the user invokes /deep-reflect or $deep-reflect. |
Egregore Deep Reflect (v4)
Native Codex Egregore skill. Ask the org's memory a question and return a
verified, cited synthesis of what it collectively knows — including what it
doesn't. /reflect captures user thought; /deep-reflect researches what the
org already knows.
Full contracts live in .claude/skills/deep-reflect/ (SKILL.md entry,
RESEARCH.md engine, ROLES.md agent roles, REPORT.md output). Follow them;
this file is the Codex-runtime adapter.
Modes
- Question (default): interrogative or bare-topic arguments → "what does the
org know about X?"
- Cross-ref: "cross-reference this" / a declarative insight → "how does this
insight sit against existing memory?"
- Depth:
--brief (1 hop wave), standard (≤3), --deep (≤5). Default is
dynamic from the corpus probe.
Flow
- Check
egregore.json mode FIRST. Local mode: full engine over files; never
call graph, batch-graph, or notification scripts; zero graph vocabulary in
output. Connected: graph adds hop rails and writeback, every graph call
non-fatal (2>/dev/null || true).
- FRAME: corpus probe — 2×
bin/agent.sh search "$TOPIC" --fast -n 20 with
different wordings + bin/artifacts.sh find "$TOPIC" +
grep -ril "$TOPIC" memory/knowledge/ memory/handoffs/ memory/artifacts/.
N = the union of unique paths; set the wave budget from N per
RESEARCH.md §Budget; decompose the question into 2-4 sub-questions;
print a plan line (Researching: … · ~N relevant docs · budget M waves · est ~X min). Under ~5 relevant docs: say memory is too thin and offer
/reflect or /harvest instead of faking research. If search errors or
returns nothing while grep hits, rely on grep and note "recall degraded".
- SEED: check
memory/knowledge/research/ for prior runs on the topic —
their cited artifacts (frontmatter artifacts_cited, older reports use
artifacts_consulted; resolve filename stems with find) are pre-warmed
leads, and their null_regions gaps get re-audited. Build a frontier of
≤12 docs from search probes, frontmatter topic/quest greps, and index
ledgers.
- HOP WAVES: the Codex runtime has no Workflow tool — run the foreground
staged path from
ROLES.md: per wave, read 8-12 docs (subagents if the
session offers them, otherwise staged single-model passes), extract
claims-with-verbatim-excerpts and typed leads (doc links, quests, people,
sessions, topic terms, named absences). Dedup against the seen-set, score
leads, hop until saturation per RESEARCH.md (coverage on every
sub-question plus low novelty or zero yield; or budget).
- PATTERN: two de-correlated passes. Without subagents true blindness is
impossible — degrade deliberately: run the TENSION pass FIRST (grouped by
topic), then the convergence pass (chronological), so tension-finding is
never anchored by an existing convergence story. Disagreement between
passes is a finding, never averaged away.
- VERIFY: for every claim, whitespace-normalize the excerpt and re-find it in
the cited file (miss = miscite = kill; supported paraphrase = correct the
excerpt). The mechanical re-check survives self-verification — it is a
grep, not a judgment. Re-attack every gap claim with one extra probe; a
found doc refutes the gap and joins the ledger. Evidence older than ~90
days gets one
--fast newer-artifact probe. Report the kill count. When
the verifying context is the same one that made the claims, mark the
report's provenance verification: self-checked.
- SYNTHESIZE + PRESENT: stance, not summary — answer first, then ◆ primary /
◇ secondary findings with
memory/... path citations, tensions, and "what
memory doesn't hold" (gaps + the probes that prove them). Confidence is
derived from evidence structure (verified + ≥3 independent sources = high;
single-source always labeled). Register: analyst, not critic.
- CAPTURE (user-gated): Save (default) / Edit first / File follow-ups /
Skip. On Save write
memory/knowledge/research/YYYY-MM-DD-author-slug.md
with frontmatter (question, mode, stop_reason, waves, docs_read,
claims_killed, builds_on, artifacts_consulted, artifacts_cited,
null_regions), then connected-mode best-effort graph writes per the main
REPORT.md §Writeback (MERGE with filePath — do NOT use
graph-op.sh register-artifact; RELATES_TO/TENSION_WITH for verified
cross-ref claims only), then
bin/agent.sh save --message "Deep research: $TOPIC" --topic "$TOPIC".
On Skip: nothing is written except the local run ledger under
.egregore/research-runs/. Render the TUI box AFTER the gate, reflecting
the actual outcome.
Output
Render the Egregore deep-reflect confirmation TUI after the capture gate:
- 72-column outer box, standard top/separator/content/bottom lines only.
- Header:
◈ DEEP REFLECT, author, date.
- Body: question, ◆/◇ finding counts, confidence, waves · docs read · cited ·
claims cut in verification.
- Footer: actual saved/pushed state; say "graphed" only if the graph write
succeeded; in local mode omit graph language entirely.
- Structured UX parity is required: preserve the rendered deep-reflect TUI
box, no preamble, no prose-only replacement, and no raw search or graph
output.
- Never show raw graph JSON; never replace the box with prose.
Rules
- Search budget: at most ONE non-
--fast search probe per run (seed only);
--fast everywhere else. Count result blocks, never the banner hit count.
- No claim without a verbatim excerpt that survives re-finding in the file.
- Absence is a first-class finding — cite the probes that prove it.
- Don't manufacture significance: "memory holds nothing structural on this"
is a valid report.
- At most ONE clarifying question before the run, plus the capture gate.
- Do not use Claude Code commands.
