Start hereTry itAgentsArchitectureMechanismsResultsGlossaryRelated workSpikeChangelogContributingSecurity

Production runs, 2026-10-12

Evidence for the README's results table, from the production deployment https://firetrident.mattsolt.workers.dev. Every landing named here is a real commit on an Artifacts trunk repository.

Redacted in every file: Artifacts tokens, bearer and API tokens, the account id in Artifacts hostnames (ACCOUNT_ID), email addresses, and absolute local paths ($TMPDIR, <scratch>, or repository-relative). Checked by scanning every file for those patterns and for the deployment's real token values. byo-transcript.jsonl is trimmed (see its row). The colony summaries were read with GET requests only, sent without a token.

File What it shows Produced by Date (UTC) Colony
hono-run.log Hono (583 files): the tsconfig prerequisite lands as a human intent, the rule finds 958 violations (282 intents), 80 LLM agents (4 on kimi-k2.7-code, 76 on gemma-4-26b) reach 0 violations in 63.1 s (polled every 5 s; 58.0 s on the colony's clock); 282 rule intents land in 192 commits, 0 returned, 0 abandoned; agent model spend $0.0524; then the jsr.json follow-up lands in 37.7 s ruby scripts/demo_hono.rb --agents 80 --model '@cf/moonshotai/kimi-k2.7-code=4,@cf/google/gemma-4-26b-a4b-it' --payoff, copied with node packages/bench/src/evidence.ts redact 2026-10-12 09:08–09:11 https://firetrident.mattsolt.workers.dev/c/hono
hono-events.jsonl The colony's own event stream during that swarm: 192 landing events carrying 282 landed proposals, no proposal returned; violations fall from 955 after the first landing to 0 at 09:10:22.941 ft watch hono --json --only proposal,landing,violations, started by scripts/demo_hono.rb; copied with evidence.ts redact 2026-10-12 09:09–09:10 https://firetrident.mattsolt.workers.dev/c/hono
shortlinks-run.log A new project grown from demo/new-project-spec.md into an empty trunk: 5 rules, 13 intents, 6 LLM agents; all 13 intents land 479.1 s after ft new in 13 commits, 0 returned, 0 rejected, 0 escalated; model spend $0.2522; from a plain git clone, node --test passes 60 of 60 and the reference acceptance suite (demo/shortlinks-reference/test/acceptance.test.ts) 7 of 7, with 0 violations of the 5 rules ruby demo/run_new_project.rb --out <dir>, copied with evidence.ts redact 2026-10-12 09:11–09:19 https://firetrident.mattsolt.workers.dev/c/shortlinks
shortlinks-events.jsonl The colony's event stream for that run: each of the 13 intents claimed, proposed and landed, 13 proposals landed in 13 landing events, none returned ft watch shortlinks --json --only proposal,landing,intent, started by demo/run_new_project.rb; copied with evidence.ts redact 2026-10-12 09:12–09:19 https://firetrident.mattsolt.workers.dev/c/shortlinks
shortlinks-review-run.log A second run of the same spec, 15 hours later, with escalations left for a person: at 152 s the deterministic danger patterns sent the authentication route (Routes: GET and DELETE /api/links/:slug) to a person; Matt approved it about 12 minutes later (ft review), and all 13 intents landed in 1,428.7 s; model spend $0.2344. From a plain clone this run is worse than the first: node --test passes 62 of 67 (KvStore click counting, readJson's empty-body and size-limit handling, and DELETE without a token fail) and the reference acceptance suite 6 of 7 (DELETE answers 405 instead of 401). No test ran on the platform before these changes landed; the 5 rule checks found 0 violations ruby demo/run_new_project.rb --name shortlinks-review --stall 1500 --timeout 3600, copied with evidence.ts redact 2026-10-13 00:12–00:36 https://firetrident.mattsolt.workers.dev/c/shortlinks-review
shortlinks-review-certify-run.log Trunk certification on that second run, after it finished: ft certify --now at 22:38 EDT runs the colony's own tests on trunk e1468ea in a Dynamic Worker (267 ms): red, 62 of 67, the same 5 failures as a plain clone; the colony declares a repair intent; 2 in-platform agents (all their model calls routed to kimi-k2.7-code) land the first repair at 22:50 (48f74b8: red, 66 of 67, the router change broke one test) and the colony declares a second; the follower stops waiting at 50 min with that repair still open; a plain clone of 48f74b8 passes 66 of 67 and the reference acceptance 7 of 7 demo/run_new_project.rb --attach --certify-wait 3000 2026-10-13 02:38–03:29 shortlinks-review
shortlinks-review-certify-green.log The same colony once the second repair landed (23:45 EDT, after two abandoned attempts and one returned by the review arc's second opinion for compile errors): trunk 97b1036 certified green, 67 of 67; a plain clone passes 67 of 67 (node --test), the reference acceptance 7 of 7, and the spec's 5 rules with 0 violations; both repair intents landed demo/run_new_project.rb --attach --certify-wait 600 2026-10-13 03:47 shortlinks-review
shortlinks-review-certify-events.jsonl The colony's event stream from the first certification to green (22:38–23:46 EDT, 77 events without metrics and marks): each certification (running, red, red, green) with its counts, both repair intents opening, claimed, abandoned and landing, the returned proposal with the reviewer's reason, and the agents' status notes GET /api/colonies/shortlinks-review/events, filtered 2026-10-13 shortlinks-review
certifications-shortlinks-review.json The colony's certification results and state (GET /api/colonies/shortlinks-review/certifications): e1468ea red 62/67, 48f74b8 red 66/67, 97b1036 green 67/67 (the green mark), each with its failing tests, timing and repair intent; repairs at most 3 an hour node packages/bench/src/evidence.ts get 2026-10-13 shortlinks-review
byo-run.log An external coding agent, run headless, joins the colony over MCP, takes an intent, edits and pushes with plain git in its own workspace, proposes and waits; its change lands on deterministic proof (joined 6.3 s, took 8.6 s, proposed 31.5 s, landed 38.1 s into the session; 41.7 s, 18 turns, $0.171) while one in-platform synthetic agent lands the colony's other two intents ruby scripts/mcp_join.rb --url https://firetrident.mattsolt.workers.dev --colony byo-demo --setup --swarm 1 --transcript <file>, copied with evidence.ts redact 2026-10-12 09:25–09:26 https://firetrident.mattsolt.workers.dev/c/byo-demo
byo-transcript.jsonl That agent session, TRIMMED from 45 stream events (85 KB) to 41 lines (21 KB): the session's model and MCP tools, every tool call with its input and result, the agent's text, one permission denial and the final result; thinking, rate-limit and token-count events, ids, usage and the client's local configuration are dropped. Token values were already <redacted> when the transcript was captured node packages/bench/src/evidence.ts transcript <captured transcript> byo-transcript.jsonl 2026-10-12 09:25–09:26 https://firetrident.mattsolt.workers.dev/c/byo-demo
vaccination-openspec-w1.json The vaccination meter (GET /api/colonies/openspec-w1/vaccinate): 28 sealed bad proposals in 4 threat classes, 7 each, injected in batches of 20 and 8 at 06:36–06:39; 28 caught (21 escalated by the reflex, 7 returned by the second-opinion model), 0 missed, Wilson 95% interval 0.879–1.0. Sealed proposals cannot land node packages/bench/src/evidence.ts get /colonies/openspec-w1/vaccinate --out vaccination-openspec-w1.json fetched 2026-10-12 12:35 https://firetrident.mattsolt.workers.dev/c/openspec-w1
colony-hono.json The Hono colony now: 284 intents landed (282 from the rule, plus the prerequisite and the follow-up), 194 landing commits, 0 returned, 0 violations; 282 landed on deterministic proof and 2 on a second-opinion model; 81 agents; the lineage note of src/index.ts line 17 node packages/bench/src/evidence.ts colony hono --why src/index.ts:17 --out colony-hono.json fetched 2026-10-12 12:36 https://firetrident.mattsolt.workers.dev/c/hono
colony-shortlinks.json The shortlinks colony now: 13 of 13 intents in 13 landing commits, 0 returned, 0 violations of the 5 rules, each landing approved by the reflex and a second-opinion model; 6 agents; the lineage of src/index.ts node packages/bench/src/evidence.ts colony shortlinks --why src/index.ts --out colony-shortlinks.json fetched 2026-10-12 12:36 https://firetrident.mattsolt.workers.dev/c/shortlinks
colony-byo-demo.json The byo-demo colony now: 3 of 3 intents in 3 landing commits, all on deterministic proof; 1 external agent and 1 synthetic in-platform agent; the lineage of benchmark.js (the external agent's landing: model, read and write sets, checks, what it did not verify) node packages/bench/src/evidence.ts colony byo-demo --why benchmark.js --out colony-byo-demo.json fetched 2026-10-12 12:36 https://firetrident.mattsolt.workers.dev/c/byo-demo
colony-openspec-e2e.json The OpenSpec colony (1,287 files, rule "Node built-ins use node:") now: 74 of 74 intents in 43 landing commits, 0 returned, 0 violations; 21 agents (20 synthetic in-platform, 1 scripted external). It predates the review layer, so it records no decisions node packages/bench/src/evidence.ts colony openspec-e2e --out colony-openspec-e2e.json fetched 2026-10-12 12:36 https://firetrident.mattsolt.workers.dev/c/openspec-e2e
colony-openspec-w1.json The review-layer colony (OpenSpec, same rule) now: 74 rule intents in 48 landing commits; funnel reflex 104, LLM 46, human 1; one human rejection minted antibody ab_mv4vopdt6oslmv, committed to trunk at .firetrident/antibodies/ab_mv4vopdt6oslmv.json; 1 escalated proposal still waits for a person; vaccination 28 of 28 node packages/bench/src/evidence.ts colony openspec-w1 --out colony-openspec-w1.json fetched 2026-10-12 12:36 https://firetrident.mattsolt.workers.dev/c/openspec-w1
colony-bench-w2-cascade.json The cascade colony behind ../2026-10-12/cascade.json now: 32 intents landed from 62 proposals, 30 returned stale, 7 landing commits node packages/bench/src/evidence.ts colony bench-w2-cascade --out colony-bench-w2-cascade.json fetched 2026-10-12 12:36 https://firetrident.mattsolt.workers.dev/c/bench-w2-cascade
colony-hono-spend-2026-10-19.json The Hono colony's metrics with its model spend as the colony prices the tokens it counted, at Workers AI list prices of 2026-10-12: $0.0648 in all ($0.0639 agents, $0.0009 review) for 244,003 input and 19,815 output tokens. It covers the whole colony as it stands: 285 landed intents in 195 landing commits and 82 agents, against 284, 194 and 81 in colony-hono.json FT_URL=https://firetrident.com node packages/bench/src/evidence.ts get /colonies/hono/metrics --out colony-hono-spend-2026-10-19.json fetched 2026-10-19 15:43 https://firetrident.com/c/hono
openspec-e2e-start-count-2026-10-19.log The OpenSpec colony's starting count, recounted: the colony imported Fission-AI/OpenSpec at 9111a76 (the base of its first landed proposal, and main's head when it was created), 1,287 files; the colony's own scan (scripts/rule_scan.ts: scanFiles and core checkFile, run on the code of 6c0a42a, as the scan stood when the colony was created; the core of 959a5c5 refused the rule's 258-character pattern; the cap is now 1,024 characters) finds 131 violations in 74 files, each listed with its lines; the colony's first violations event reads 131 the steps in the log: a stock clone, git archive, node scripts/rule_scan.ts, and GETs with no token 2026-10-19 15:54 https://firetrident.com/c/openspec-e2e

The colony summaries list the 20 most recent landings, because the API returns no more. The full history of each colony is its trunk repository: every landing commit carries FT-* trailers and a JSON lineage note under refs/notes/firetrident.

In this folder