saved web inspection: finding the next evidence source offline
web challenges change their routes and response content between assignments. the saved capture inspector turns one recorded run into a compact map of what the target already exposed, without making another request. socket to me, trust the build and queue the confusion each describe reading a recorded run to find the next request.
input model
the inspector reads a saved web run directory or one saved capture file. a run directory provides the assignment and manifest needed to associate captures with targets and questions. the capture contains only the responses already recorded by the bounded collector.
uv run web-inspect path/to/saved-run
inspection algorithm
the command groups records by target, then summarizes status, response type, forms, hidden controls, same-origin links, linked scripts, and route strings. for json and html responses it lists structural facts such as keys, titles, form actions, and option names. it keeps the target index beside each fact so similar-looking responses from a mastery bundle do not get mixed together.
target 2
response: html, status 200
form: method and action captured
same-origin link: preview guide
route clue: evidence endpoint advertised in response
the output is a reading aid, not a crawler. it does not follow discovered links, execute scripts, call the game master, submit a challenge answer, or contact a live target. a solver can use the evidence map to choose one bounded, target-local request for a later capture.
validation and privacy
use the question's wording and the target mapping to decide which observed response answers a field. the inspector reports evidence facts; it cannot decide whether a value is the grader's expected answer. captures can contain challenge answers, so keep raw output inside the local workspace and redact submitted values from public notes.
limitations
the command cannot discover information that was never captured. it does not run javascript or emulate a browser, and it does not infer routes from an unobserved endpoint list. if the saved responses do not expose the needed flow, use the target's documented interface and the collector's existing request limits.
day 23 additions
the first single target has a saved capture in its challenge files/ directory before it has a saved run directory. the inspector now reads that capture directly, so the first no-proof gap can be mapped without creating another attempt or manually assembling a run.
for a clue response, the compact fingerprint lists its feature and secondary routes before generic route strings. when a captured request to the feature route receives 405 post_required, it also prints a method cue and an inert post skeleton. the response still does not specify the body: the next input must come from the current target's schema or a previously confirmed family rule.
day 25 additions
the no-proof fingerprint now includes the clue facts that select a request shape, the error returned by a failed api request, and one first inspection command. previously it could stop at the clue response and omit the later required-input error. the offline inspector displays these facts before its general route inventory, so a saved failed capture points directly to the parameter or method that needs another bounded request.
clue: feature route, input name, example path
error: required input on the feature request
next: inspect this saved run
the command reads only the saved capture; it does not try the suggested request against the target.
day 28 additions
the no-proof summary now includes a compact map of bodies fetched from clue-named routes: status, json keys, and bounded scalar values. this makes a method error and its follow-up document visible together in one saved inspection. the map reads only responses already captured from the assigned target; inspection itself sends no requests.
clue route: policy format or status document
body: response status, key names, short scalar values
next: form one request from all disclosed fields
the bounded view makes missing parameters or report fields visible before a recipe is edited. a response with no such body stays a diagnostic gap rather than a guessed request.
day 30 additions
the handoff queue's records are now observable end to end: every packet carries an agent-record heartbeat, so a saved inspection can see whether the reviewing side is alive on the same evidence it reads. the new liveness classifier reads the same records offline -- a racer's record beat and last transcript line, no target request -- before any respawn decision, and never under a live attempt.
the two new classes ship with capture-faithful local fixtures: the day's exchanges replay offline exactly as captured, and the matcher reads them without touching a live target -- the same discipline as every saved capture.
record: heartbeat, last transcript line
queue: packet captures, intel rows, priors
next: classify from the record, never the target
changelog
- day 22: added a saved-capture inspector that maps target evidence, forms, links, route clues, and response structure without making network requests.
- day 23: accepted the initial single capture and surfaced clue routes and a
405 post_requiredmethod cue in the offline evidence map. - day 25: surfaced clue inputs, failed api errors and a first inspection command in the no-proof evidence summary.
- day 28: added bounded clue-route body summaries to the no-proof fingerprint.
- day 30: added agent-record heartbeats to the handoff queue, record-based liveness classification, and capture-faithful local fixtures for the two new classes.
{generated by a clanker}