Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
git clone --depth 1 https://github.com/NicoGenti/opencode-orchestrator-kitWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/commands/nicogenti/opencode-orchestrator-kit/recall)<a href="https://agentmods.dev/commands/nicogenti/opencode-orchestrator-kit/recall"><img src="https://agentmods.dev/badge/commands/nicogenti/opencode-orchestrator-kit/recall/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/commands/nicogenti/opencode-orchestrator-kit/recall"><img src="https://agentmods.dev/badge/commands/nicogenti/opencode-orchestrator-kit/recall.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00023 | $0.01575 |
| Opus 5.5 | $0.00009 | $0.00630 |
| Sonnet 5.5 | $0.00005 | $0.00315 |
| Haiku 4.5 | $0.00002 | $0.00158 |
Grade A, and why
recall scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 52 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Comando /recall. Flusso di ri-attivazione di una sessione di comprensione passata per un plan. È un flow strettamente read-only di diagnostica e ri-verifica della comprensione — mai una replay del lavoro, mai una re-delegazione implicita.
Invariant: ALL questions are shown BEFORE any code or explanation is displayed.
Pre-condition: .context/comprehension-log.md is read ONLY during explicit /recall invocation — never at session start, never during bootstrap.
Flow (9 steps):
- List candidates: scan
plan/complete/*.mdand existing.context/comprehension/**/*.mdrecords; present to user as a numbered list (plan-id + evaluation_id + research_id + last outcome + date). Each numbered entry is ONE evaluation — selecting it targets exactly oneevaluation_idand its storedresearch_id; the completed recall appends a recall event, it never rewrites history. - User selects the evaluation to recall (number or evaluation_id); the plan-id is used only if that plan has exactly ONE evaluation.
- Load questions: read the per-evaluation record at
.context/comprehension/<evaluation-id>.mdif it exists (src=stored; the user's selection pins one specificresearch_id); otherwise reconstruct from the plan's Goal/Scope section inplan/complete/<plan-id>.md(src=reconstructed), inferring questions that map to the plan's stated acceptance criteria. - Show ALL questions (with
- ?prefix), one per line. No code, no diffs, no explanations yet. - User answers from memory (no tooling, no file access during this step); from answer time on, reads go only through step 5's whitelist — nothing else.
- Code inspection (v0.3.1 token-aware comprehension, v0.6.1 caps): if needed, show the relevant diff with
maxDiffLines=300and correlated symbols capped atmaxRelatedSymbols=3— diff lines ONLY, NEVER full files, NEVER repo-wide scans. - Evaluation delegated to
comprehension-coachusing the v0.4.0 DCI rubric with the identical whitelist input (goal, changed file list, focused diff, minimal surrounding symbols). The coach's rubric score is the DCI verdict. - Retry/skip (v0.3.0 verbatim): maximum 1 retry;
skip comprehensionis always available; no reveal of expected answers during retry. - Outcome: append dci1-format row to
.context/comprehension-log.mdAND full-overwrite.context/progress.md(snapshot of the CURRENT state — pointer removal on completion, never a per-task history line); the immediate baseline (dci_immediate, user_confidence_immediate, outcome_immediate) is immutable — never rewritten; a new recall event is a NEW line — no in-place rewrite of the record or the dataset (v0.6.1+ normalisation). Format:
<date> | plan=<NNNN> | type=dci1 | src=stored|reconstructed | retry=<0|1> | user_conf=<1-5>|n/a | evaluator_conf=<1-5>|n/a | DCI=<score>/<available_score> | DCI=skipped | outcome=PASS|FAIL|RETRY|SKIPPED
-
outcome=SKIPPEDrow format:user_conf=n/a | evaluator_conf=n/a | DCI=skipped | outcome=SKIPPED(no numeric confidence fields). -
v0.6.1: the two confidence fields are distinct normalised values —
user_confis the developer's calibration self-rating,evaluator_confthe coach's confidence in its verdict; the old overloaded singleconf=field is retired. -
Research dataset event (v0.6.6): after the dci1 evaluation,
/recallAPPENDS one recall event — a NEW line — to.context/research-dataset.jsonl, carrying the SAMEresearch_idrecorded by the task event (stored in the per-evaluation record at task end; the record supplies the id — the dataset itself is never edited or rewritten to find it). If the record carries noresearch_id(task predates research mode), nothing is written to the dataset. No modify, no backfill, no delete, no rewrite of existing rows — the dataset is a strictly append-only event log. Never inside a task lifecycle. Theresearch_idformat isres-YYYYMMDD-HHMMSS-<32hex>(v0.6.6: 128-bit random CSPRNG suffix):/recallreads it from the per-evaluation record and copies it verbatim — never regenerates, reformats, or truncates it. -
Same header-once rule as the telemetry writer; same append-only discipline.
-
The log file
.context/comprehension-log.mdis the single sink for all dci1 rows.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- yesterday First seen · 52 lines · 23 tokens per session scan A 3c317548a133
recall is a command published in the GitHub repository NicoGenti/opencode-orchestrator-kit (8 stars, last pushed yesterday), licensed MIT. It adds 23 tokens to every session and 1,575 once invoked, about $0.0001 per session on Opus 5.5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-10-27.
Other commands, from other repositories
quiz
Check your understanding of what was built — recall questions grounded in the session log, with results tracked across sessions.
learn
Show Claude Code best practices and save confirmed session learnings.
explain
Guided code tour of a file or subsystem this session touched — entry point, the load-bearing pieces, the edges, and what connects to it.
moodle-new-plugin
You will scaffold a new Moodle plugin.
aic-help
Show AIContext quickstart guide with setup, workflows, and best practices.
aigon-help
Quick-allow when prompted: Claude Shift+Tab • Antigravity uses --dangerously-skip-permissions at launch • Cursor "Add to allowlist" • Codex "Allow and remember".